diff options
Diffstat (limited to 'src/core/lib')
55 files changed, 2011 insertions, 681 deletions
diff --git a/src/core/lib/channel/channel_args.c b/src/core/lib/channel/channel_args.c index 3a56b1ff20..2957d2c818 100644 --- a/src/core/lib/channel/channel_args.c +++ b/src/core/lib/channel/channel_args.c @@ -272,6 +272,18 @@ int grpc_channel_args_compare(const grpc_channel_args *a, return 0; } +const grpc_arg *grpc_channel_args_find(const grpc_channel_args *args, + const char *name) { + if (args != NULL) { + for (size_t i = 0; i < args->num_args; ++i) { + if (strcmp(args->args[i].key, name) == 0) { + return &args->args[i]; + } + } + } + return NULL; +} + int grpc_channel_arg_get_integer(grpc_arg *arg, grpc_integer_options options) { if (arg->type != GRPC_ARG_INTEGER) { gpr_log(GPR_ERROR, "%s ignored: it must be an integer", arg->key); diff --git a/src/core/lib/channel/channel_args.h b/src/core/lib/channel/channel_args.h index 586a296d1f..a80340c0fa 100644 --- a/src/core/lib/channel/channel_args.h +++ b/src/core/lib/channel/channel_args.h @@ -89,6 +89,10 @@ uint32_t grpc_channel_args_compression_algorithm_get_states( int grpc_channel_args_compare(const grpc_channel_args *a, const grpc_channel_args *b); +/** Returns the value of argument \a name from \a args, or NULL if not found. */ +const grpc_arg *grpc_channel_args_find(const grpc_channel_args *args, + const char *name); + typedef struct grpc_integer_options { int default_value; // Return this if value is outside of expected bounds. int min_value; diff --git a/src/core/lib/channel/channel_stack.c b/src/core/lib/channel/channel_stack.c index 57d34d9e9a..2c5367901d 100644 --- a/src/core/lib/channel/channel_stack.c +++ b/src/core/lib/channel/channel_stack.c @@ -162,7 +162,7 @@ grpc_error *grpc_call_stack_init( grpc_exec_ctx *exec_ctx, grpc_channel_stack *channel_stack, int initial_refs, grpc_iomgr_cb_func destroy, void *destroy_arg, grpc_call_context_element *context, const void *transport_server_data, - gpr_timespec deadline, grpc_call_stack *call_stack) { + grpc_mdstr *path, gpr_timespec deadline, grpc_call_stack *call_stack) { grpc_channel_element *channel_elems = CHANNEL_ELEMS_FROM_STACK(channel_stack); grpc_call_element_args args; size_t count = channel_stack->count; @@ -179,10 +179,12 @@ grpc_error *grpc_call_stack_init( /* init per-filter data */ grpc_error *first_error = GRPC_ERROR_NONE; + args.start_time = gpr_now(GPR_CLOCK_MONOTONIC); for (i = 0; i < count; i++) { args.call_stack = call_stack; args.server_transport_data = transport_server_data; args.context = context; + args.path = path; args.deadline = deadline; call_elems[i].filter = channel_elems[i].filter; call_elems[i].channel_data = channel_elems[i].channel_data; diff --git a/src/core/lib/channel/channel_stack.h b/src/core/lib/channel/channel_stack.h index 1cfe2885d8..27f3be7b29 100644 --- a/src/core/lib/channel/channel_stack.h +++ b/src/core/lib/channel/channel_stack.h @@ -74,6 +74,8 @@ typedef struct { grpc_call_stack *call_stack; const void *server_transport_data; grpc_call_context_element *context; + grpc_mdstr *path; + gpr_timespec start_time; gpr_timespec deadline; } grpc_call_element_args; @@ -225,7 +227,7 @@ grpc_error *grpc_call_stack_init( grpc_exec_ctx *exec_ctx, grpc_channel_stack *channel_stack, int initial_refs, grpc_iomgr_cb_func destroy, void *destroy_arg, grpc_call_context_element *context, const void *transport_server_data, - gpr_timespec deadline, grpc_call_stack *call_stack); + grpc_mdstr *path, gpr_timespec deadline, grpc_call_stack *call_stack); /* Set a pollset or a pollset_set for a call stack: must occur before the first * op is started */ void grpc_call_stack_set_pollset_or_pollset_set(grpc_exec_ctx *exec_ctx, diff --git a/src/core/lib/channel/deadline_filter.c b/src/core/lib/channel/deadline_filter.c index 079b98a2f8..d2ea5250f6 100644 --- a/src/core/lib/channel/deadline_filter.c +++ b/src/core/lib/channel/deadline_filter.c @@ -64,30 +64,49 @@ static void timer_callback(grpc_exec_ctx* exec_ctx, void* arg, } // Starts the deadline timer. -static void start_timer_if_needed(grpc_exec_ctx* exec_ctx, - grpc_call_element* elem, - gpr_timespec deadline) { +static void start_timer_if_needed_locked(grpc_exec_ctx* exec_ctx, + grpc_call_element* elem, + gpr_timespec deadline) { grpc_deadline_state* deadline_state = elem->call_data; deadline = gpr_convert_clock_type(deadline, GPR_CLOCK_MONOTONIC); - if (gpr_time_cmp(deadline, gpr_inf_future(GPR_CLOCK_MONOTONIC)) != 0) { + // Note: We do not start the timer if there is already a timer + // pending. This should be okay, because this is only called from two + // functions exported by this module: grpc_deadline_state_start(), which + // starts the initial timer, and grpc_deadline_state_reset(), which + // cancels any pre-existing timer before starting a new one. In + // particular, we want to ensure that if grpc_deadline_state_start() + // winds up trying to start the timer after grpc_deadline_state_reset() + // has already done so, we ignore the value from the former. + if (!deadline_state->timer_pending && + gpr_time_cmp(deadline, gpr_inf_future(GPR_CLOCK_MONOTONIC)) != 0) { // Take a reference to the call stack, to be owned by the timer. GRPC_CALL_STACK_REF(deadline_state->call_stack, "deadline_timer"); - gpr_mu_lock(&deadline_state->timer_mu); deadline_state->timer_pending = true; grpc_timer_init(exec_ctx, &deadline_state->timer, deadline, timer_callback, elem, gpr_now(GPR_CLOCK_MONOTONIC)); - gpr_mu_unlock(&deadline_state->timer_mu); } } +static void start_timer_if_needed(grpc_exec_ctx* exec_ctx, + grpc_call_element* elem, + gpr_timespec deadline) { + grpc_deadline_state* deadline_state = elem->call_data; + gpr_mu_lock(&deadline_state->timer_mu); + start_timer_if_needed_locked(exec_ctx, elem, deadline); + gpr_mu_unlock(&deadline_state->timer_mu); +} // Cancels the deadline timer. -static void cancel_timer_if_needed(grpc_exec_ctx* exec_ctx, - grpc_deadline_state* deadline_state) { - gpr_mu_lock(&deadline_state->timer_mu); +static void cancel_timer_if_needed_locked(grpc_exec_ctx* exec_ctx, + grpc_deadline_state* deadline_state) { if (deadline_state->timer_pending) { grpc_timer_cancel(exec_ctx, &deadline_state->timer); deadline_state->timer_pending = false; } +} +static void cancel_timer_if_needed(grpc_exec_ctx* exec_ctx, + grpc_deadline_state* deadline_state) { + gpr_mu_lock(&deadline_state->timer_mu); + cancel_timer_if_needed_locked(exec_ctx, deadline_state); gpr_mu_unlock(&deadline_state->timer_mu); } @@ -108,6 +127,21 @@ static void inject_on_complete_cb(grpc_deadline_state* deadline_state, op->on_complete = &deadline_state->on_complete; } +void grpc_deadline_state_init(grpc_exec_ctx* exec_ctx, grpc_call_element* elem, + grpc_call_stack* call_stack) { + grpc_deadline_state* deadline_state = elem->call_data; + memset(deadline_state, 0, sizeof(*deadline_state)); + deadline_state->call_stack = call_stack; + gpr_mu_init(&deadline_state->timer_mu); +} + +void grpc_deadline_state_destroy(grpc_exec_ctx* exec_ctx, + grpc_call_element* elem) { + grpc_deadline_state* deadline_state = elem->call_data; + cancel_timer_if_needed(exec_ctx, deadline_state); + gpr_mu_destroy(&deadline_state->timer_mu); +} + // Callback and associated state for starting the timer after call stack // initialization has been completed. struct start_timer_after_init_state { @@ -122,16 +156,11 @@ static void start_timer_after_init(grpc_exec_ctx* exec_ctx, void* arg, gpr_free(state); } -void grpc_deadline_state_init(grpc_exec_ctx* exec_ctx, grpc_call_element* elem, - grpc_call_element_args* args) { - grpc_deadline_state* deadline_state = elem->call_data; - memset(deadline_state, 0, sizeof(*deadline_state)); - deadline_state->call_stack = args->call_stack; - gpr_mu_init(&deadline_state->timer_mu); +void grpc_deadline_state_start(grpc_exec_ctx* exec_ctx, grpc_call_element* elem, + gpr_timespec deadline) { // Deadline will always be infinite on servers, so the timer will only be // set on clients with a finite deadline. - const gpr_timespec deadline = - gpr_convert_clock_type(args->deadline, GPR_CLOCK_MONOTONIC); + deadline = gpr_convert_clock_type(deadline, GPR_CLOCK_MONOTONIC); if (gpr_time_cmp(deadline, gpr_inf_future(GPR_CLOCK_MONOTONIC)) != 0) { // When the deadline passes, we indicate the failure by sending down // an op with cancel_error set. However, we can't send down any ops @@ -148,11 +177,13 @@ void grpc_deadline_state_init(grpc_exec_ctx* exec_ctx, grpc_call_element* elem, } } -void grpc_deadline_state_destroy(grpc_exec_ctx* exec_ctx, - grpc_call_element* elem) { +void grpc_deadline_state_reset(grpc_exec_ctx* exec_ctx, grpc_call_element* elem, + gpr_timespec new_deadline) { grpc_deadline_state* deadline_state = elem->call_data; - cancel_timer_if_needed(exec_ctx, deadline_state); - gpr_mu_destroy(&deadline_state->timer_mu); + gpr_mu_lock(&deadline_state->timer_mu); + cancel_timer_if_needed_locked(exec_ctx, deadline_state); + start_timer_if_needed_locked(exec_ctx, elem, new_deadline); + gpr_mu_unlock(&deadline_state->timer_mu); } void grpc_deadline_state_client_start_transport_stream_op( @@ -209,7 +240,8 @@ static grpc_error* init_call_elem(grpc_exec_ctx* exec_ctx, grpc_call_element_args* args) { // Note: size of call data is different between client and server. memset(elem->call_data, 0, elem->filter->sizeof_call_data); - grpc_deadline_state_init(exec_ctx, elem, args); + grpc_deadline_state_init(exec_ctx, elem, args->call_stack); + grpc_deadline_state_start(exec_ctx, elem, args->deadline); return GRPC_ERROR_NONE; } diff --git a/src/core/lib/channel/deadline_filter.h b/src/core/lib/channel/deadline_filter.h index 685df87761..716a852565 100644 --- a/src/core/lib/channel/deadline_filter.h +++ b/src/core/lib/channel/deadline_filter.h @@ -54,18 +54,37 @@ typedef struct grpc_deadline_state { grpc_closure* next_on_complete; } grpc_deadline_state; -// To be used in a filter's init_call_elem(), destroy_call_elem(), and -// start_transport_stream_op() methods to enforce call deadlines. // -// REQUIRES: The first field in elem->call_data is a grpc_deadline_state. +// NOTE: All of these functions require that the first field in +// elem->call_data is a grpc_deadline_state. // -// For grpc_deadline_state_client_start_transport_stream_op(), it is the -// caller's responsibility to chain to the next filter if necessary -// after the function returns. + void grpc_deadline_state_init(grpc_exec_ctx* exec_ctx, grpc_call_element* elem, - grpc_call_element_args* args); + grpc_call_stack* call_stack); void grpc_deadline_state_destroy(grpc_exec_ctx* exec_ctx, grpc_call_element* elem); + +// Starts the timer with the specified deadline. +// Should be called from the filter's init_call_elem() method. +void grpc_deadline_state_start(grpc_exec_ctx* exec_ctx, grpc_call_element* elem, + gpr_timespec deadline); + +// Cancels the existing timer and starts a new one with new_deadline. +// +// Note: It is generally safe to call this with an earlier deadline +// value than the current one, but not the reverse. No checks are done +// to ensure that the timer callback is not invoked while it is in the +// process of being reset, which means that attempting to increase the +// deadline may result in the timer being called twice. +void grpc_deadline_state_reset(grpc_exec_ctx* exec_ctx, grpc_call_element* elem, + gpr_timespec new_deadline); + +// To be called from the client-side filter's start_transport_stream_op() +// method. Ensures that the deadline timer is cancelled when the call +// is completed. +// +// Note: It is the caller's responsibility to chain to the next filter if +// necessary after this function returns. void grpc_deadline_state_client_start_transport_stream_op( grpc_exec_ctx* exec_ctx, grpc_call_element* elem, grpc_transport_stream_op* op); diff --git a/src/core/lib/channel/handshaker.c b/src/core/lib/channel/handshaker.c index 8f9fb17a31..0d759887bc 100644 --- a/src/core/lib/channel/handshaker.c +++ b/src/core/lib/channel/handshaker.c @@ -183,7 +183,7 @@ void grpc_handshake_manager_do_handshake( gpr_timespec deadline, grpc_tcp_server_acceptor* acceptor, grpc_handshaker_done_cb cb, void* user_data) { grpc_channel_args* args_copy = grpc_channel_args_copy(args); - gpr_slice_buffer* read_buffer = malloc(sizeof(*read_buffer)); + gpr_slice_buffer* read_buffer = gpr_malloc(sizeof(*read_buffer)); gpr_slice_buffer_init(read_buffer); if (mgr->count == 0) { // No handshakers registered, so we just immediately call the done diff --git a/src/core/lib/channel/http_server_filter.c b/src/core/lib/channel/http_server_filter.c index 0f2bf97824..f2221fb0fb 100644 --- a/src/core/lib/channel/http_server_filter.c +++ b/src/core/lib/channel/http_server_filter.c @@ -42,6 +42,8 @@ #define EXPECTED_CONTENT_TYPE "application/grpc" #define EXPECTED_CONTENT_TYPE_LENGTH sizeof(EXPECTED_CONTENT_TYPE) - 1 +extern int grpc_http_trace; + typedef struct call_data { uint8_t seen_path; uint8_t seen_method; @@ -209,6 +211,11 @@ static void hs_on_recv(grpc_exec_ctx *exec_ctx, void *user_data, err, GRPC_ERROR_CREATE("Missing te: trailers header")); } /* Error this call out */ + if (grpc_http_trace) { + const char *error_str = grpc_error_string(err); + gpr_log(GPR_ERROR, "Invalid http2 headers: %s", error_str); + grpc_error_free_string(error_str); + } grpc_call_element_send_cancel(exec_ctx, elem); } } else { diff --git a/src/core/lib/channel/message_size_filter.c b/src/core/lib/channel/message_size_filter.c index f067a3a51c..b8b2546035 100644 --- a/src/core/lib/channel/message_size_filter.c +++ b/src/core/lib/channel/message_size_filter.c @@ -38,13 +38,54 @@ #include <grpc/support/log.h> #include <grpc/support/string_util.h> +#include "src/core/ext/client_channel/method_config.h" #include "src/core/lib/channel/channel_args.h" #define DEFAULT_MAX_SEND_MESSAGE_LENGTH -1 // Unlimited. // The protobuf library will (by default) start warning at 100 megs. #define DEFAULT_MAX_RECV_MESSAGE_LENGTH (4 * 1024 * 1024) +typedef struct message_size_limits { + int max_send_size; + int max_recv_size; +} message_size_limits; + +static void* message_size_limits_copy(void* value) { + void* new_value = gpr_malloc(sizeof(message_size_limits)); + memcpy(new_value, value, sizeof(message_size_limits)); + return new_value; +} + +static int message_size_limits_cmp(void* value1, void* value2) { + const message_size_limits* v1 = value1; + const message_size_limits* v2 = value2; + if (v1->max_send_size > v2->max_send_size) return 1; + if (v1->max_send_size < v2->max_send_size) return -1; + if (v1->max_recv_size > v2->max_recv_size) return 1; + if (v1->max_recv_size < v2->max_recv_size) return -1; + return 0; +} + +static const grpc_mdstr_hash_table_vtable message_size_limits_vtable = { + gpr_free, message_size_limits_copy, message_size_limits_cmp}; + +static void* method_config_convert_value( + const grpc_method_config* method_config) { + message_size_limits* value = gpr_malloc(sizeof(message_size_limits)); + const int32_t* max_request_message_bytes = + grpc_method_config_get_max_request_message_bytes(method_config); + value->max_send_size = + max_request_message_bytes != NULL ? *max_request_message_bytes : -1; + const int32_t* max_response_message_bytes = + grpc_method_config_get_max_response_message_bytes(method_config); + value->max_recv_size = + max_response_message_bytes != NULL ? *max_response_message_bytes : -1; + return value; +} + typedef struct call_data { + int max_send_size; + int max_recv_size; // Receive closures are chained: we inject this closure as the // recv_message_ready up-call on transport_stream_op, and remember to // call our next_recv_message_ready member after handling it. @@ -58,6 +99,8 @@ typedef struct call_data { typedef struct channel_data { int max_send_size; int max_recv_size; + // Maps path names to message_size_limits structs. + grpc_mdstr_hash_table* method_limit_table; } channel_data; // Callback invoked when we receive a message. Here we check the max @@ -66,13 +109,12 @@ static void recv_message_ready(grpc_exec_ctx* exec_ctx, void* user_data, grpc_error* error) { grpc_call_element* elem = user_data; call_data* calld = elem->call_data; - channel_data* chand = elem->channel_data; - if (*calld->recv_message != NULL && chand->max_recv_size >= 0 && - (*calld->recv_message)->length > (size_t)chand->max_recv_size) { + if (*calld->recv_message != NULL && calld->max_recv_size >= 0 && + (*calld->recv_message)->length > (size_t)calld->max_recv_size) { char* message_string; gpr_asprintf(&message_string, "Received message larger than max (%u vs. %d)", - (*calld->recv_message)->length, chand->max_recv_size); + (*calld->recv_message)->length, calld->max_recv_size); grpc_error* new_error = grpc_error_set_int( GRPC_ERROR_CREATE(message_string), GRPC_ERROR_INT_GRPC_STATUS, GRPC_STATUS_INVALID_ARGUMENT); @@ -93,13 +135,12 @@ static void start_transport_stream_op(grpc_exec_ctx* exec_ctx, grpc_call_element* elem, grpc_transport_stream_op* op) { call_data* calld = elem->call_data; - channel_data* chand = elem->channel_data; // Check max send message size. - if (op->send_message != NULL && chand->max_send_size >= 0 && - op->send_message->length > (size_t)chand->max_send_size) { + if (op->send_message != NULL && calld->max_send_size >= 0 && + op->send_message->length > (size_t)calld->max_send_size) { char* message_string; gpr_asprintf(&message_string, "Sent message larger than max (%u vs. %d)", - op->send_message->length, chand->max_send_size); + op->send_message->length, calld->max_send_size); gpr_slice message = gpr_slice_from_copied_string(message_string); gpr_free(message_string); grpc_call_element_send_close_with_message( @@ -119,9 +160,32 @@ static void start_transport_stream_op(grpc_exec_ctx* exec_ctx, static grpc_error* init_call_elem(grpc_exec_ctx* exec_ctx, grpc_call_element* elem, grpc_call_element_args* args) { + channel_data* chand = elem->channel_data; call_data* calld = elem->call_data; calld->next_recv_message_ready = NULL; grpc_closure_init(&calld->recv_message_ready, recv_message_ready, elem); + // Get max sizes from channel data, then merge in per-method config values. + // Note: Per-method config is only available on the client, so we + // apply the max request size to the send limit and the max response + // size to the receive limit. + calld->max_send_size = chand->max_send_size; + calld->max_recv_size = chand->max_recv_size; + if (chand->method_limit_table != NULL) { + message_size_limits* limits = + grpc_method_config_table_get(chand->method_limit_table, args->path); + if (limits != NULL) { + if (limits->max_send_size >= 0 && + (limits->max_send_size < calld->max_send_size || + calld->max_send_size < 0)) { + calld->max_send_size = limits->max_send_size; + } + if (limits->max_recv_size >= 0 && + (limits->max_recv_size < calld->max_recv_size || + calld->max_recv_size < 0)) { + calld->max_recv_size = limits->max_recv_size; + } + } + } return GRPC_ERROR_NONE; } @@ -155,11 +219,23 @@ static void init_channel_elem(grpc_exec_ctx* exec_ctx, grpc_channel_arg_get_integer(&args->channel_args->args[i], options); } } + // Get method config table from channel args. + const grpc_arg* channel_arg = + grpc_channel_args_find(args->channel_args, GRPC_ARG_SERVICE_CONFIG); + if (channel_arg != NULL) { + GPR_ASSERT(channel_arg->type == GRPC_ARG_POINTER); + chand->method_limit_table = grpc_method_config_table_convert( + (grpc_method_config_table*)channel_arg->value.pointer.p, + method_config_convert_value, &message_size_limits_vtable); + } } // Destructor for channel_data. static void destroy_channel_elem(grpc_exec_ctx* exec_ctx, - grpc_channel_element* elem) {} + grpc_channel_element* elem) { + channel_data* chand = elem->channel_data; + grpc_mdstr_hash_table_unref(chand->method_limit_table); +} const grpc_channel_filter grpc_message_size_filter = { start_transport_stream_op, diff --git a/src/core/lib/iomgr/closure.c b/src/core/lib/iomgr/closure.c index 1ba0a5c141..c6ddc76732 100644 --- a/src/core/lib/iomgr/closure.c +++ b/src/core/lib/iomgr/closure.c @@ -35,6 +35,8 @@ #include <grpc/support/alloc.h> +#include "src/core/lib/profiling/timers.h" + void grpc_closure_init(grpc_closure *closure, grpc_iomgr_cb_func cb, void *cb_arg) { closure->cb = cb; @@ -51,7 +53,7 @@ void grpc_closure_list_append(grpc_closure_list *closure_list, GRPC_ERROR_UNREF(error); return; } - closure->error = error; + closure->error_data.error = error; closure->next_data.next = NULL; if (closure_list->head == NULL) { closure_list->head = closure; @@ -64,8 +66,8 @@ void grpc_closure_list_append(grpc_closure_list *closure_list, void grpc_closure_list_fail_all(grpc_closure_list *list, grpc_error *forced_failure) { for (grpc_closure *c = list->head; c != NULL; c = c->next_data.next) { - if (c->error == GRPC_ERROR_NONE) { - c->error = GRPC_ERROR_REF(forced_failure); + if (c->error_data.error == GRPC_ERROR_NONE) { + c->error_data.error = GRPC_ERROR_REF(forced_failure); } } GRPC_ERROR_UNREF(forced_failure); @@ -110,3 +112,13 @@ grpc_closure *grpc_closure_create(grpc_iomgr_cb_func cb, void *cb_arg) { grpc_closure_init(&wc->wrapper, closure_wrapper, wc); return &wc->wrapper; } + +void grpc_closure_run(grpc_exec_ctx *exec_ctx, grpc_closure *c, + grpc_error *error) { + GPR_TIMER_BEGIN("grpc_closure_run", 0); + if (c != NULL) { + c->cb(exec_ctx, c->cb_arg, error); + } + GRPC_ERROR_UNREF(error); + GPR_TIMER_END("grpc_closure_run", 0); +} diff --git a/src/core/lib/iomgr/closure.h b/src/core/lib/iomgr/closure.h index c1a22b6021..2b4b271eaa 100644 --- a/src/core/lib/iomgr/closure.h +++ b/src/core/lib/iomgr/closure.h @@ -76,7 +76,10 @@ struct grpc_closure { void *cb_arg; /** Once queued, the result of the closure. Before then: scratch space */ - grpc_error *error; + union { + grpc_error *error; + uintptr_t scratch; + } error_data; }; /** Initializes \a closure with \a cb and \a cb_arg. */ @@ -106,4 +109,10 @@ void grpc_closure_list_move(grpc_closure_list *src, grpc_closure_list *dst); /** return whether \a list is empty. */ bool grpc_closure_list_empty(grpc_closure_list list); +/** Run a closure directly. Caller ensures that no locks are being held above. + * Note that calling this at the end of a closure callback function itself is + * by definition safe. */ +void grpc_closure_run(grpc_exec_ctx *exec_ctx, grpc_closure *closure, + grpc_error *error); + #endif /* GRPC_CORE_LIB_IOMGR_CLOSURE_H */ diff --git a/src/core/lib/iomgr/combiner.c b/src/core/lib/iomgr/combiner.c index 831bdb4aff..60ee14eb23 100644 --- a/src/core/lib/iomgr/combiner.c +++ b/src/core/lib/iomgr/combiner.c @@ -50,25 +50,57 @@ int grpc_combiner_trace = 0; } \ } while (0) +#define STATE_UNORPHANED 1 +#define STATE_ELEM_COUNT_LOW_BIT 2 + struct grpc_combiner { + grpc_combiner *next_combiner_on_this_exec_ctx; grpc_workqueue *optional_workqueue; gpr_mpscq queue; // state is: - // lower bit - zero if orphaned - // other bits - number of items queued on the lock + // lower bit - zero if orphaned (STATE_UNORPHANED) + // other bits - number of items queued on the lock (STATE_ELEM_COUNT_LOW_BIT) gpr_atm state; - bool take_async_break_before_final_list; + // number of elements in the list that are covered by a poller: if >0, we can + // offload safely + gpr_atm elements_covered_by_poller; + bool time_to_execute_final_list; + bool final_list_covered_by_poller; grpc_closure_list final_list; - grpc_closure continue_finishing; + grpc_closure offload; }; +static void offload(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error); + +typedef struct { + grpc_error *error; + bool covered_by_poller; +} error_data; + +static uintptr_t pack_error_data(error_data d) { + return ((uintptr_t)d.error) | (d.covered_by_poller ? 1 : 0); +} + +static error_data unpack_error_data(uintptr_t p) { + return (error_data){(grpc_error *)(p & ~(uintptr_t)1), p & 1}; +} + +static bool is_covered_by_poller(grpc_combiner *lock) { + return lock->final_list_covered_by_poller || + gpr_atm_acq_load(&lock->elements_covered_by_poller) > 0; +} + grpc_combiner *grpc_combiner_create(grpc_workqueue *optional_workqueue) { grpc_combiner *lock = gpr_malloc(sizeof(*lock)); + lock->next_combiner_on_this_exec_ctx = NULL; + lock->time_to_execute_final_list = false; lock->optional_workqueue = optional_workqueue; - gpr_atm_no_barrier_store(&lock->state, 1); + lock->final_list_covered_by_poller = false; + gpr_atm_no_barrier_store(&lock->state, STATE_UNORPHANED); + gpr_atm_no_barrier_store(&lock->elements_covered_by_poller, 0); gpr_mpscq_init(&lock->queue); - lock->take_async_break_before_final_list = false; grpc_closure_list_init(&lock->final_list); + grpc_closure_init(&lock->offload, offload, lock); GRPC_COMBINER_TRACE(gpr_log(GPR_DEBUG, "C:%p create", lock)); return lock; } @@ -82,7 +114,7 @@ static void really_destroy(grpc_exec_ctx *exec_ctx, grpc_combiner *lock) { } void grpc_combiner_destroy(grpc_exec_ctx *exec_ctx, grpc_combiner *lock) { - gpr_atm old_state = gpr_atm_full_fetch_add(&lock->state, -1); + gpr_atm old_state = gpr_atm_full_fetch_add(&lock->state, -STATE_UNORPHANED); GRPC_COMBINER_TRACE(gpr_log( GPR_DEBUG, "C:%p really_destroy old_state=%" PRIdPTR, lock, old_state)); if (old_state == 1) { @@ -90,170 +122,186 @@ void grpc_combiner_destroy(grpc_exec_ctx *exec_ctx, grpc_combiner *lock) { } } -static bool maybe_finish_one(grpc_exec_ctx *exec_ctx, grpc_combiner *lock); -static void finish(grpc_exec_ctx *exec_ctx, grpc_combiner *lock); +static void push_last_on_exec_ctx(grpc_exec_ctx *exec_ctx, + grpc_combiner *lock) { + lock->next_combiner_on_this_exec_ctx = NULL; + if (exec_ctx->active_combiner == NULL) { + exec_ctx->active_combiner = exec_ctx->last_combiner = lock; + } else { + exec_ctx->last_combiner->next_combiner_on_this_exec_ctx = lock; + exec_ctx->last_combiner = lock; + } +} -static void continue_finishing_mainline(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error) { - GPR_TIMER_BEGIN("combiner.continue_executing_mainline", 0); - grpc_combiner *lock = arg; - GRPC_COMBINER_TRACE( - gpr_log(GPR_DEBUG, "C:%p continue_finishing_mainline", lock)); - GPR_ASSERT(exec_ctx->active_combiner == NULL); +static void push_first_on_exec_ctx(grpc_exec_ctx *exec_ctx, + grpc_combiner *lock) { + lock->next_combiner_on_this_exec_ctx = exec_ctx->active_combiner; exec_ctx->active_combiner = lock; - if (maybe_finish_one(exec_ctx, lock)) finish(exec_ctx, lock); - GPR_ASSERT(exec_ctx->active_combiner == lock); - exec_ctx->active_combiner = NULL; - GPR_TIMER_END("combiner.continue_executing_mainline", 0); + if (lock->next_combiner_on_this_exec_ctx == NULL) { + exec_ctx->last_combiner = lock; + } } -static void execute_final(grpc_exec_ctx *exec_ctx, grpc_combiner *lock) { - GPR_TIMER_BEGIN("combiner.execute_final", 0); - grpc_closure *c = lock->final_list.head; - GPR_ASSERT(c != NULL); - grpc_closure_list_init(&lock->final_list); - lock->take_async_break_before_final_list = false; - int loops = 0; - while (c != NULL) { - GRPC_COMBINER_TRACE( - gpr_log(GPR_DEBUG, "C:%p execute_final[%d] c=%p", lock, loops, c)); - grpc_closure *next = c->next_data.next; - grpc_error *error = c->error; - c->cb(exec_ctx, c->cb_arg, error); - GRPC_ERROR_UNREF(error); - c = next; - loops++; +void grpc_combiner_execute(grpc_exec_ctx *exec_ctx, grpc_combiner *lock, + grpc_closure *cl, grpc_error *error, + bool covered_by_poller) { + GPR_TIMER_BEGIN("combiner.execute", 0); + gpr_atm last = gpr_atm_full_fetch_add(&lock->state, STATE_ELEM_COUNT_LOW_BIT); + GRPC_COMBINER_TRACE(gpr_log( + GPR_DEBUG, "C:%p grpc_combiner_execute c=%p cov=%d last=%" PRIdPTR, lock, + cl, covered_by_poller, last)); + GPR_ASSERT(last & STATE_UNORPHANED); // ensure lock has not been destroyed + cl->error_data.scratch = + pack_error_data((error_data){error, covered_by_poller}); + if (covered_by_poller) { + gpr_atm_no_barrier_fetch_add(&lock->elements_covered_by_poller, 1); + } + gpr_mpscq_push(&lock->queue, &cl->next_data.atm_next); + if (last == 1) { + // first element on this list: add it to the list of combiner locks + // executing within this exec_ctx + push_last_on_exec_ctx(exec_ctx, lock); } - GPR_TIMER_END("combiner.execute_final", 0); + GPR_TIMER_END("combiner.execute", 0); } -static void continue_executing_final(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error) { - GPR_TIMER_BEGIN("combiner.continue_executing_final", 0); - grpc_combiner *lock = arg; - GRPC_COMBINER_TRACE( - gpr_log(GPR_DEBUG, "C:%p continue_executing_final", lock)); - GPR_ASSERT(exec_ctx->active_combiner == NULL); - exec_ctx->active_combiner = lock; - // quick peek to see if new things have turned up on the queue: if so, go back - // to executing them before the final list - if ((gpr_atm_acq_load(&lock->state) >> 1) > 1) { - if (maybe_finish_one(exec_ctx, lock)) finish(exec_ctx, lock); - } else { - execute_final(exec_ctx, lock); - finish(exec_ctx, lock); +static void move_next(grpc_exec_ctx *exec_ctx) { + exec_ctx->active_combiner = + exec_ctx->active_combiner->next_combiner_on_this_exec_ctx; + if (exec_ctx->active_combiner == NULL) { + exec_ctx->last_combiner = NULL; } - GPR_ASSERT(exec_ctx->active_combiner == lock); - exec_ctx->active_combiner = NULL; - GPR_TIMER_END("combiner.continue_executing_final", 0); } -static bool start_execute_final(grpc_exec_ctx *exec_ctx, grpc_combiner *lock) { - GPR_TIMER_BEGIN("combiner.start_execute_final", 0); - GPR_ASSERT(exec_ctx->active_combiner == lock); - GRPC_COMBINER_TRACE( - gpr_log(GPR_DEBUG, - "C:%p start_execute_final take_async_break_before_final_list=%d", - lock, lock->take_async_break_before_final_list)); - if (lock->take_async_break_before_final_list) { - grpc_closure_init(&lock->continue_finishing, continue_executing_final, - lock); - grpc_exec_ctx_sched(exec_ctx, &lock->continue_finishing, GRPC_ERROR_NONE, - GRPC_WORKQUEUE_REF(lock->optional_workqueue, "sched")); - GPR_TIMER_END("combiner.start_execute_final", 0); +static void offload(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { + grpc_combiner *lock = arg; + push_last_on_exec_ctx(exec_ctx, lock); +} + +static void queue_offload(grpc_exec_ctx *exec_ctx, grpc_combiner *lock) { + move_next(exec_ctx); + GRPC_COMBINER_TRACE(gpr_log(GPR_DEBUG, "C:%p queue_offload --> %p", lock, + lock->optional_workqueue)); + grpc_workqueue_enqueue(exec_ctx, lock->optional_workqueue, &lock->offload, + GRPC_ERROR_NONE); +} + +bool grpc_combiner_continue_exec_ctx(grpc_exec_ctx *exec_ctx) { + GPR_TIMER_BEGIN("combiner.continue_exec_ctx", 0); + grpc_combiner *lock = exec_ctx->active_combiner; + if (lock == NULL) { + GPR_TIMER_END("combiner.continue_exec_ctx", 0); return false; - } else { - execute_final(exec_ctx, lock); - GPR_TIMER_END("combiner.start_execute_final", 0); - return true; } -} -static bool maybe_finish_one(grpc_exec_ctx *exec_ctx, grpc_combiner *lock) { - GPR_TIMER_BEGIN("combiner.maybe_finish_one", 0); - gpr_mpscq_node *n = gpr_mpscq_pop(&lock->queue); GRPC_COMBINER_TRACE( - gpr_log(GPR_DEBUG, "C:%p maybe_finish_one n=%p", lock, n)); - GPR_ASSERT(exec_ctx->active_combiner == lock); - if (n == NULL) { - // Queue is in an transiently inconsistent state: a new item is being queued - // but is not visible to this thread yet. - // Use this as a cue that we should go off and do something else for a while - // (and come back later) - grpc_closure_init(&lock->continue_finishing, continue_finishing_mainline, - lock); - grpc_exec_ctx_sched(exec_ctx, &lock->continue_finishing, GRPC_ERROR_NONE, - GRPC_WORKQUEUE_REF(lock->optional_workqueue, "sched")); - GPR_TIMER_END("combiner.maybe_finish_one", 0); - return false; + gpr_log(GPR_DEBUG, + "C:%p grpc_combiner_continue_exec_ctx workqueue=%p " + "is_covered_by_poller=%d exec_ctx_ready_to_finish=%d " + "time_to_execute_final_list=%d", + lock, lock->optional_workqueue, is_covered_by_poller(lock), + grpc_exec_ctx_ready_to_finish(exec_ctx), + lock->time_to_execute_final_list)); + + if (lock->optional_workqueue != NULL && is_covered_by_poller(lock) && + grpc_exec_ctx_ready_to_finish(exec_ctx)) { + GPR_TIMER_MARK("offload_from_finished_exec_ctx", 0); + // this execution context wants to move on, and we have a workqueue (and + // so can help the execution context out): schedule remaining work to be + // picked up on the workqueue + queue_offload(exec_ctx, lock); + GPR_TIMER_END("combiner.continue_exec_ctx", 0); + return true; } - grpc_closure *cl = (grpc_closure *)n; - grpc_error *error = cl->error; - cl->cb(exec_ctx, cl->cb_arg, error); - GRPC_ERROR_UNREF(error); - GPR_TIMER_END("combiner.maybe_finish_one", 0); - return true; -} -static void finish(grpc_exec_ctx *exec_ctx, grpc_combiner *lock) { - bool (*executor)(grpc_exec_ctx * exec_ctx, grpc_combiner * lock); - GPR_TIMER_BEGIN("combiner.finish", 0); - int loops = 0; - do { - executor = maybe_finish_one; - gpr_atm old_state = gpr_atm_full_fetch_add(&lock->state, -2); - GRPC_COMBINER_TRACE(gpr_log(GPR_DEBUG, - "C:%p finish[%d] old_state=%" PRIdPTR, lock, - loops, old_state)); - switch (old_state) { - default: - // we have multiple queued work items: just continue executing them - break; - case 5: // we're down to one queued item: if it's the final list we - case 4: // should do that - if (!grpc_closure_list_empty(lock->final_list)) { - executor = start_execute_final; - } - break; - case 3: // had one count, one unorphaned --> unlocked unorphaned - GPR_TIMER_END("combiner.finish", 0); - return; - case 2: // and one count, one orphaned --> unlocked and orphaned - really_destroy(exec_ctx, lock); - GPR_TIMER_END("combiner.finish", 0); - return; - case 1: - case 0: - // these values are illegal - representing an already unlocked or - // deleted lock - GPR_UNREACHABLE_CODE(return ); + if (!lock->time_to_execute_final_list || + // peek to see if something new has shown up, and execute that with + // priority + (gpr_atm_acq_load(&lock->state) >> 1) > 1) { + gpr_mpscq_node *n = gpr_mpscq_pop(&lock->queue); + GRPC_COMBINER_TRACE( + gpr_log(GPR_DEBUG, "C:%p maybe_finish_one n=%p", lock, n)); + if (n == NULL) { + // queue is in an inconsistent state: use this as a cue that we should + // go off and do something else for a while (and come back later) + GPR_TIMER_MARK("delay_busy", 0); + if (lock->optional_workqueue != NULL && is_covered_by_poller(lock)) { + queue_offload(exec_ctx, lock); + } + GPR_TIMER_END("combiner.continue_exec_ctx", 0); + return true; } - loops++; - } while (executor(exec_ctx, lock)); - GPR_TIMER_END("combiner.finish", 0); -} + GPR_TIMER_BEGIN("combiner.exec1", 0); + grpc_closure *cl = (grpc_closure *)n; + error_data err = unpack_error_data(cl->error_data.scratch); + cl->cb(exec_ctx, cl->cb_arg, err.error); + if (err.covered_by_poller) { + gpr_atm_no_barrier_fetch_add(&lock->elements_covered_by_poller, -1); + } + GRPC_ERROR_UNREF(err.error); + GPR_TIMER_END("combiner.exec1", 0); + } else { + grpc_closure *c = lock->final_list.head; + GPR_ASSERT(c != NULL); + grpc_closure_list_init(&lock->final_list); + lock->final_list_covered_by_poller = false; + int loops = 0; + while (c != NULL) { + GPR_TIMER_BEGIN("combiner.exec_1final", 0); + GRPC_COMBINER_TRACE( + gpr_log(GPR_DEBUG, "C:%p execute_final[%d] c=%p", lock, loops, c)); + grpc_closure *next = c->next_data.next; + grpc_error *error = c->error_data.error; + c->cb(exec_ctx, c->cb_arg, error); + GRPC_ERROR_UNREF(error); + c = next; + GPR_TIMER_END("combiner.exec_1final", 0); + } + } -void grpc_combiner_execute(grpc_exec_ctx *exec_ctx, grpc_combiner *lock, - grpc_closure *cl, grpc_error *error) { + GPR_TIMER_MARK("unref", 0); + move_next(exec_ctx); + lock->time_to_execute_final_list = false; + gpr_atm old_state = + gpr_atm_full_fetch_add(&lock->state, -STATE_ELEM_COUNT_LOW_BIT); GRPC_COMBINER_TRACE( - gpr_log(GPR_DEBUG, "C:%p grpc_combiner_execute c=%p", lock, cl)); - GPR_TIMER_BEGIN("combiner.execute", 0); - gpr_atm last = gpr_atm_full_fetch_add(&lock->state, 2); - GPR_ASSERT(last & 1); // ensure lock has not been destroyed - if (last == 1) { - exec_ctx->active_combiner = lock; - GPR_TIMER_BEGIN("combiner.execute_first_cb", 0); - cl->cb(exec_ctx, cl->cb_arg, error); - GPR_TIMER_END("combiner.execute_first_cb", 0); - GRPC_ERROR_UNREF(error); - finish(exec_ctx, lock); - GPR_ASSERT(exec_ctx->active_combiner == lock); - exec_ctx->active_combiner = NULL; - } else { - cl->error = error; - gpr_mpscq_push(&lock->queue, &cl->next_data.atm_next); + gpr_log(GPR_DEBUG, "C:%p finish old_state=%" PRIdPTR, lock, old_state)); +// Define a macro to ease readability of the following switch statement. +#define OLD_STATE_WAS(orphaned, elem_count) \ + (((orphaned) ? 0 : STATE_UNORPHANED) | \ + ((elem_count)*STATE_ELEM_COUNT_LOW_BIT)) + // Depending on what the previous state was, we need to perform different + // actions. + switch (old_state) { + default: + // we have multiple queued work items: just continue executing them + break; + case OLD_STATE_WAS(false, 2): + case OLD_STATE_WAS(true, 2): + // we're down to one queued item: if it's the final list we should do that + if (!grpc_closure_list_empty(lock->final_list)) { + lock->time_to_execute_final_list = true; + } + break; + case OLD_STATE_WAS(false, 1): + // had one count, one unorphaned --> unlocked unorphaned + GPR_TIMER_END("combiner.continue_exec_ctx", 0); + return true; + case OLD_STATE_WAS(true, 1): + // and one count, one orphaned --> unlocked and orphaned + really_destroy(exec_ctx, lock); + GPR_TIMER_END("combiner.continue_exec_ctx", 0); + return true; + case OLD_STATE_WAS(false, 0): + case OLD_STATE_WAS(true, 0): + // these values are illegal - representing an already unlocked or + // deleted lock + GPR_TIMER_END("combiner.continue_exec_ctx", 0); + GPR_UNREACHABLE_CODE(return true); } - GPR_TIMER_END("combiner.execute", 0); + push_first_on_exec_ctx(exec_ctx, lock); + GPR_TIMER_END("combiner.continue_exec_ctx", 0); + return true; } static void enqueue_finally(grpc_exec_ctx *exec_ctx, void *closure, @@ -264,30 +312,26 @@ static void enqueue_finally(grpc_exec_ctx *exec_ctx, void *closure, void grpc_combiner_execute_finally(grpc_exec_ctx *exec_ctx, grpc_combiner *lock, grpc_closure *closure, grpc_error *error, - bool force_async_break) { + bool covered_by_poller) { GRPC_COMBINER_TRACE(gpr_log( - GPR_DEBUG, - "C:%p grpc_combiner_execute_finally c=%p force_async_break=%d; ac=%p", - lock, closure, force_async_break, exec_ctx->active_combiner)); + GPR_DEBUG, "C:%p grpc_combiner_execute_finally c=%p; ac=%p; cov=%d", lock, + closure, exec_ctx->active_combiner, covered_by_poller)); GPR_TIMER_BEGIN("combiner.execute_finally", 0); if (exec_ctx->active_combiner != lock) { GPR_TIMER_MARK("slowpath", 0); grpc_combiner_execute(exec_ctx, lock, - grpc_closure_create(enqueue_finally, closure), error); + grpc_closure_create(enqueue_finally, closure), error, + false); GPR_TIMER_END("combiner.execute_finally", 0); return; } - if (force_async_break) { - lock->take_async_break_before_final_list = true; - } if (grpc_closure_list_empty(lock->final_list)) { - gpr_atm_full_fetch_add(&lock->state, 2); + gpr_atm_full_fetch_add(&lock->state, STATE_ELEM_COUNT_LOW_BIT); + } + if (covered_by_poller) { + lock->final_list_covered_by_poller = true; } grpc_closure_list_append(&lock->final_list, closure, error); GPR_TIMER_END("combiner.execute_finally", 0); } - -void grpc_combiner_force_async_finally(grpc_combiner *lock) { - lock->take_async_break_before_final_list = true; -} diff --git a/src/core/lib/iomgr/combiner.h b/src/core/lib/iomgr/combiner.h index 1409db24b9..d04eeed83a 100644 --- a/src/core/lib/iomgr/combiner.h +++ b/src/core/lib/iomgr/combiner.h @@ -52,19 +52,14 @@ grpc_combiner *grpc_combiner_create(grpc_workqueue *optional_workqueue); void grpc_combiner_destroy(grpc_exec_ctx *exec_ctx, grpc_combiner *lock); // Execute \a action within the lock. void grpc_combiner_execute(grpc_exec_ctx *exec_ctx, grpc_combiner *lock, - grpc_closure *closure, grpc_error *error); + grpc_closure *closure, grpc_error *error, + bool covered_by_poller); // Execute \a action within the lock just prior to unlocking. -// if \a hint_async_break is true, the combiner tries to hand execution to -// another thread before finishing the primary queue of combined closures and -// executing the finally list. -// Deprecation warning: \a hint_async_break will be removed in a future version -// Takes a very slow and round-about path if not called from a -// grpc_combiner_execute closure. void grpc_combiner_execute_finally(grpc_exec_ctx *exec_ctx, grpc_combiner *lock, grpc_closure *closure, grpc_error *error, - bool hint_async_break); -// Deprecated: force the finally list execution onto another thread -void grpc_combiner_force_async_finally(grpc_combiner *lock); + bool covered_by_poller); + +bool grpc_combiner_continue_exec_ctx(grpc_exec_ctx *exec_ctx); extern int grpc_combiner_trace; diff --git a/src/core/lib/iomgr/error.c b/src/core/lib/iomgr/error.c index 31c80260f8..f6bb3a0477 100644 --- a/src/core/lib/iomgr/error.c +++ b/src/core/lib/iomgr/error.c @@ -120,6 +120,8 @@ static const char *error_int_name(grpc_error_ints key) { return "http_status"; case GRPC_ERROR_INT_LIMIT: return "limit"; + case GRPC_ERROR_INT_OCCURRED_DURING_WRITE: + return "occurred_during_write"; } GPR_UNREACHABLE_CODE(return "unknown"); } @@ -144,6 +146,8 @@ static const char *error_str_name(grpc_error_strs key) { return "tsi_error"; case GRPC_ERROR_STR_FILENAME: return "filename"; + case GRPC_ERROR_STR_QUEUED_BUFFERS: + return "queued_buffers"; } GPR_UNREACHABLE_CODE(return "unknown"); } @@ -265,7 +269,7 @@ static grpc_error *copy_error_and_unref(grpc_error *in) { } else { out = gpr_malloc(sizeof(*out)); #ifdef GRPC_ERROR_REFCOUNT_DEBUG - gpr_log(GPR_DEBUG, "%p create copying", out); + gpr_log(GPR_DEBUG, "%p create copying %p", out, in); #endif out->ints = gpr_avl_ref(in->ints); out->strs = gpr_avl_ref(in->strs); @@ -523,21 +527,25 @@ static char *fmt_time(void *p) { return out; } -static void add_errs(gpr_avl_node *n, char **s, size_t *sz, size_t *cap) { +static void add_errs(gpr_avl_node *n, char **s, size_t *sz, size_t *cap, + bool *first) { if (n == NULL) return; - add_errs(n->left, s, sz, cap); + add_errs(n->left, s, sz, cap, first); + if (!*first) append_chr(',', s, sz, cap); + *first = false; const char *e = grpc_error_string(n->value); append_str(e, s, sz, cap); grpc_error_free_string(e); - add_errs(n->right, s, sz, cap); + add_errs(n->right, s, sz, cap, first); } static char *errs_string(grpc_error *err) { char *s = NULL; size_t sz = 0; size_t cap = 0; + bool first = true; append_chr('[', &s, &sz, &cap); - add_errs(err->errs.root, &s, &sz, &cap); + add_errs(err->errs.root, &s, &sz, &cap, &first); append_chr(']', &s, &sz, &cap); append_chr(0, &s, &sz, &cap); return s; diff --git a/src/core/lib/iomgr/error.h b/src/core/lib/iomgr/error.h index 00ace8a7a9..f3f3b80a09 100644 --- a/src/core/lib/iomgr/error.h +++ b/src/core/lib/iomgr/error.h @@ -100,6 +100,8 @@ typedef enum { GRPC_ERROR_INT_HTTP_STATUS, /// context sensitive limit associated with the error GRPC_ERROR_INT_LIMIT, + /// chttp2: did the error occur while a write was in progress + GRPC_ERROR_INT_OCCURRED_DURING_WRITE, } grpc_error_ints; typedef enum { @@ -121,6 +123,8 @@ typedef enum { GRPC_ERROR_STR_TSI_ERROR, /// filename that we were trying to read/write when this error occurred GRPC_ERROR_STR_FILENAME, + /// which data was queued for writing when the error occurred + GRPC_ERROR_STR_QUEUED_BUFFERS } grpc_error_strs; typedef enum { @@ -128,9 +132,13 @@ typedef enum { GRPC_ERROR_TIME_CREATED, } grpc_error_times; +/// The following "special" errors can be propagated without allocating memory. +/// They are always even so that other code (particularly combiner locks) can +/// safely use the lower bit for themselves. + #define GRPC_ERROR_NONE ((grpc_error *)NULL) -#define GRPC_ERROR_OOM ((grpc_error *)1) -#define GRPC_ERROR_CANCELLED ((grpc_error *)2) +#define GRPC_ERROR_OOM ((grpc_error *)2) +#define GRPC_ERROR_CANCELLED ((grpc_error *)4) const char *grpc_error_string(grpc_error *error); void grpc_error_free_string(const char *str); diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c index ab77ebc78b..e5909d9380 100644 --- a/src/core/lib/iomgr/ev_epoll_linux.c +++ b/src/core/lib/iomgr/ev_epoll_linux.c @@ -152,20 +152,20 @@ static void fd_global_shutdown(void); * Polling island Declarations */ -//#define GRPC_PI_REF_COUNT_DEBUG -#ifdef GRPC_PI_REF_COUNT_DEBUG +#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG #define PI_ADD_REF(p, r) pi_add_ref_dbg((p), (r), __FILE__, __LINE__) #define PI_UNREF(exec_ctx, p, r) \ pi_unref_dbg((exec_ctx), (p), (r), __FILE__, __LINE__) -#else /* defined(GRPC_PI_REF_COUNT_DEBUG) */ +#else /* defined(GRPC_WORKQUEUE_REFCOUNT_DEBUG) */ #define PI_ADD_REF(p, r) pi_add_ref((p)) #define PI_UNREF(exec_ctx, p, r) pi_unref((exec_ctx), (p)) #endif /* !defined(GPRC_PI_REF_COUNT_DEBUG) */ +/* This is also used as grpc_workqueue (by directly casing it) */ typedef struct polling_island { gpr_mu mu; /* Ref count. Use PI_ADD_REF() and PI_UNREF() macros to increment/decrement @@ -185,8 +185,17 @@ typedef struct polling_island { * (except mu and ref_count) are invalid and must be ignored. */ gpr_atm merged_to; - /* The workqueue associated with this polling island */ - grpc_workqueue *workqueue; + /* Number of threads currently polling on this island */ + gpr_atm poller_count; + /* Mutex guarding the read end of the workqueue (must be held to pop from + * workqueue_items) */ + gpr_mu workqueue_read_mu; + /* Queue of closures to be executed */ + gpr_mpscq workqueue_items; + /* Count of items in workqueue_items */ + gpr_atm workqueue_item_count; + /* Wakeup fd used to wake pollers to check the contents of workqueue_items */ + grpc_wakeup_fd workqueue_wakeup_fd; /* The fd of the underlying epoll set */ int epoll_fd; @@ -275,6 +284,10 @@ static bool append_error(grpc_error **composite, grpc_error *error, threads that woke up MUST NOT call grpc_wakeup_fd_consume_wakeup() */ static grpc_wakeup_fd polling_island_wakeup_fd; +/* The polling island being polled right now. + See comments in workqueue_maybe_wakeup for why this is tracked. */ +static __thread polling_island *g_current_thread_polling_island; + /* Forward declaration */ static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi); @@ -289,12 +302,12 @@ static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi); gpr_atm g_epoll_sync; #endif /* defined(GRPC_TSAN) */ -#ifdef GRPC_PI_REF_COUNT_DEBUG static void pi_add_ref(polling_island *pi); static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi); -static void pi_add_ref_dbg(polling_island *pi, char *reason, char *file, - int line) { +#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG +static void pi_add_ref_dbg(polling_island *pi, const char *reason, + const char *file, int line) { long old_cnt = gpr_atm_acq_load(&pi->ref_count); pi_add_ref(pi); gpr_log(GPR_DEBUG, "Add ref pi: %p, old: %ld -> new:%ld (%s) - (%s, %d)", @@ -302,12 +315,42 @@ static void pi_add_ref_dbg(polling_island *pi, char *reason, char *file, } static void pi_unref_dbg(grpc_exec_ctx *exec_ctx, polling_island *pi, - char *reason, char *file, int line) { + const char *reason, const char *file, int line) { long old_cnt = gpr_atm_acq_load(&pi->ref_count); pi_unref(exec_ctx, pi); gpr_log(GPR_DEBUG, "Unref pi: %p, old:%ld -> new:%ld (%s) - (%s, %d)", (void *)pi, old_cnt, (old_cnt - 1), reason, file, line); } + +static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue, + const char *file, int line, + const char *reason) { + if (workqueue != NULL) { + pi_add_ref_dbg((polling_island *)workqueue, reason, file, line); + } + return workqueue; +} + +static void workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, + const char *file, int line, const char *reason) { + if (workqueue != NULL) { + pi_unref_dbg(exec_ctx, (polling_island *)workqueue, reason, file, line); + } +} +#else +static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue) { + if (workqueue != NULL) { + pi_add_ref((polling_island *)workqueue); + } + return workqueue; +} + +static void workqueue_unref(grpc_exec_ctx *exec_ctx, + grpc_workqueue *workqueue) { + if (workqueue != NULL) { + pi_unref(exec_ctx, (polling_island *)workqueue); + } +} #endif static void pi_add_ref(polling_island *pi) { @@ -315,10 +358,7 @@ static void pi_add_ref(polling_island *pi) { } static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi) { - /* If ref count went to one, we're back to just the workqueue owning a ref. - Unref the workqueue to break the loop. - - If ref count went to zero, delete the polling island. + /* If ref count went to zero, delete the polling island. Note that this deletion not be done under a lock. Once the ref count goes to zero, we are guaranteed that no one else holds a reference to the polling island (and that there is no racing pi_add_ref() call either). @@ -326,20 +366,12 @@ static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi) { Also, if we are deleting the polling island and the merged_to field is non-empty, we should remove a ref to the merged_to polling island */ - switch (gpr_atm_full_fetch_add(&pi->ref_count, -1)) { - case 2: /* last external ref: the only one now owned is by the workqueue */ - GRPC_WORKQUEUE_UNREF(exec_ctx, pi->workqueue, "polling_island"); - break; - case 1: { - polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); - polling_island_delete(exec_ctx, pi); - if (next != NULL) { - PI_UNREF(exec_ctx, next, "pi_delete"); /* Recursive call */ - } - break; + if (1 == gpr_atm_full_fetch_add(&pi->ref_count, -1)) { + polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to); + polling_island_delete(exec_ctx, pi); + if (next != NULL) { + PI_UNREF(exec_ctx, next, "pi_delete"); /* Recursive call */ } - case 0: - GPR_UNREACHABLE_CODE(return ); } } @@ -488,11 +520,20 @@ static polling_island *polling_island_create(grpc_exec_ctx *exec_ctx, pi->fd_capacity = 0; pi->fds = NULL; pi->epoll_fd = -1; - pi->workqueue = NULL; + + gpr_mu_init(&pi->workqueue_read_mu); + gpr_mpscq_init(&pi->workqueue_items); + gpr_atm_rel_store(&pi->workqueue_item_count, 0); gpr_atm_rel_store(&pi->ref_count, 0); + gpr_atm_rel_store(&pi->poller_count, 0); gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL); + if (!append_error(error, grpc_wakeup_fd_init(&pi->workqueue_wakeup_fd), + err_desc)) { + goto done; + } + pi->epoll_fd = epoll_create1(EPOLL_CLOEXEC); if (pi->epoll_fd < 0) { @@ -501,26 +542,14 @@ static polling_island *polling_island_create(grpc_exec_ctx *exec_ctx, } polling_island_add_wakeup_fd_locked(pi, &grpc_global_wakeup_fd, error); + polling_island_add_wakeup_fd_locked(pi, &pi->workqueue_wakeup_fd, error); if (initial_fd != NULL) { polling_island_add_fds_locked(pi, &initial_fd, 1, true, error); } - if (append_error(error, grpc_workqueue_create(exec_ctx, &pi->workqueue), - err_desc) && - *error == GRPC_ERROR_NONE) { - polling_island_add_fds_locked(pi, &pi->workqueue->wakeup_read_fd, 1, true, - error); - GPR_ASSERT(pi->workqueue->wakeup_read_fd->polling_island == NULL); - pi->workqueue->wakeup_read_fd->polling_island = pi; - PI_ADD_REF(pi, "fd"); - } - done: if (*error != GRPC_ERROR_NONE) { - if (pi->workqueue != NULL) { - GRPC_WORKQUEUE_UNREF(exec_ctx, pi->workqueue, "polling_island"); - } polling_island_delete(exec_ctx, pi); pi = NULL; } @@ -533,7 +562,11 @@ static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi) { if (pi->epoll_fd >= 0) { close(pi->epoll_fd); } + GPR_ASSERT(gpr_atm_no_barrier_load(&pi->workqueue_item_count) == 0); + gpr_mu_destroy(&pi->workqueue_read_mu); + gpr_mpscq_destroy(&pi->workqueue_items); gpr_mu_destroy(&pi->mu); + grpc_wakeup_fd_destroy(&pi->workqueue_wakeup_fd); gpr_free(pi->fds); gpr_free(pi); } @@ -678,6 +711,45 @@ static void polling_island_unlock_pair(polling_island *p, polling_island *q) { } } +static void workqueue_maybe_wakeup(polling_island *pi) { + /* If this thread is the current poller, then it may be that it's about to + decrement the current poller count, so we need to look past this thread */ + bool is_current_poller = (g_current_thread_polling_island == pi); + gpr_atm min_current_pollers_for_wakeup = is_current_poller ? 1 : 0; + gpr_atm current_pollers = gpr_atm_no_barrier_load(&pi->poller_count); + /* Only issue a wakeup if it's likely that some poller could come in and take + it right now. Note that since we do an anticipatory mpscq_pop every poll + loop, it's ok if we miss the wakeup here, as we'll get the work item when + the next poller enters anyway. */ + if (current_pollers > min_current_pollers_for_wakeup) { + GRPC_LOG_IF_ERROR("workqueue_wakeup_fd", + grpc_wakeup_fd_wakeup(&pi->workqueue_wakeup_fd)); + } +} + +static void workqueue_move_items_to_parent(polling_island *q) { + polling_island *p = (polling_island *)gpr_atm_no_barrier_load(&q->merged_to); + if (p == NULL) { + return; + } + gpr_mu_lock(&q->workqueue_read_mu); + int num_added = 0; + while (gpr_atm_no_barrier_load(&q->workqueue_item_count) > 0) { + gpr_mpscq_node *n = gpr_mpscq_pop(&q->workqueue_items); + if (n != NULL) { + gpr_atm_no_barrier_fetch_add(&q->workqueue_item_count, -1); + gpr_atm_no_barrier_fetch_add(&p->workqueue_item_count, 1); + gpr_mpscq_push(&p->workqueue_items, n); + num_added++; + } + } + gpr_mu_unlock(&q->workqueue_read_mu); + if (num_added > 0) { + workqueue_maybe_wakeup(p); + } + workqueue_move_items_to_parent(p); +} + static polling_island *polling_island_merge(polling_island *p, polling_island *q, grpc_error **error) { @@ -702,6 +774,8 @@ static polling_island *polling_island_merge(polling_island *p, /* Add the 'merged_to' link from p --> q */ gpr_atm_rel_store(&p->merged_to, (gpr_atm)q); PI_ADD_REF(q, "pi_merge"); /* To account for the new incoming ref from p */ + + workqueue_move_items_to_parent(q); } /* else if p == q, nothing needs to be done */ @@ -712,6 +786,26 @@ static polling_island *polling_island_merge(polling_island *p, return q; } +static void workqueue_enqueue(grpc_exec_ctx *exec_ctx, + grpc_workqueue *workqueue, grpc_closure *closure, + grpc_error *error) { + GPR_TIMER_BEGIN("workqueue.enqueue", 0); + /* take a ref to the workqueue: otherwise it can happen that whatever events + * this kicks off ends up destroying the workqueue before this function + * completes */ + GRPC_WORKQUEUE_REF(workqueue, "enqueue"); + polling_island *pi = (polling_island *)workqueue; + gpr_atm last = gpr_atm_no_barrier_fetch_add(&pi->workqueue_item_count, 1); + closure->error_data.error = error; + gpr_mpscq_push(&pi->workqueue_items, &closure->next_data.atm_next); + if (last == 0) { + workqueue_maybe_wakeup(pi); + } + workqueue_move_items_to_parent(pi); + GRPC_WORKQUEUE_UNREF(exec_ctx, workqueue, "enqueue"); + GPR_TIMER_END("workqueue.enqueue", 0); +} + static grpc_error *polling_island_global_init() { grpc_error *error = GRPC_ERROR_NONE; @@ -1042,11 +1136,8 @@ static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, static grpc_workqueue *fd_get_workqueue(grpc_fd *fd) { gpr_mu_lock(&fd->mu); - grpc_workqueue *workqueue = NULL; - if (fd->polling_island != NULL) { - workqueue = - GRPC_WORKQUEUE_REF(fd->polling_island->workqueue, "get_workqueue"); - } + grpc_workqueue *workqueue = GRPC_WORKQUEUE_REF( + (grpc_workqueue *)fd->polling_island, "fd_get_workqueue"); gpr_mu_unlock(&fd->mu); return workqueue; } @@ -1299,7 +1390,29 @@ static void pollset_reset(grpc_pollset *pollset) { GPR_ASSERT(pollset->polling_island == NULL); } -#define GRPC_EPOLL_MAX_EVENTS 1000 +static bool maybe_do_workqueue_work(grpc_exec_ctx *exec_ctx, + polling_island *pi) { + if (gpr_mu_trylock(&pi->workqueue_read_mu)) { + gpr_mpscq_node *n = gpr_mpscq_pop(&pi->workqueue_items); + gpr_mu_unlock(&pi->workqueue_read_mu); + if (n != NULL) { + if (gpr_atm_full_fetch_add(&pi->workqueue_item_count, -1) > 1) { + workqueue_maybe_wakeup(pi); + } + grpc_closure *c = (grpc_closure *)n; + grpc_closure_run(exec_ctx, c, c->error_data.error); + return true; + } else if (gpr_atm_no_barrier_load(&pi->workqueue_item_count) > 0) { + /* n == NULL might mean there's work but it's not available to be popped + * yet - try to ensure another workqueue wakes up to check shortly if so + */ + workqueue_maybe_wakeup(pi); + } + } + return false; +} + +#define GRPC_EPOLL_MAX_EVENTS 100 /* Note: sig_mask contains the signal mask to use *during* epoll_wait() */ static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, @@ -1354,7 +1467,13 @@ static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx, PI_ADD_REF(pi, "ps_work"); gpr_mu_unlock(&pollset->mu); - do { + /* If we get some workqueue work to do, it might end up completing an item on + the completion queue, so there's no need to poll... so we skip that and + redo the complete loop to verify */ + if (!maybe_do_workqueue_work(exec_ctx, pi)) { + gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); + g_current_thread_polling_island = pi; + GRPC_SCHEDULING_START_BLOCKING_REGION; ep_rv = epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, sig_mask); @@ -1386,6 +1505,11 @@ static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx, append_error(error, grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd), err_desc); + } else if (data_ptr == &pi->workqueue_wakeup_fd) { + append_error(error, + grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd), + err_desc); + maybe_do_workqueue_work(exec_ctx, pi); } else if (data_ptr == &polling_island_wakeup_fd) { GRPC_POLLING_TRACE( "pollset_work: pollset: %p, worker: %p polling island (epoll_fd: " @@ -1408,7 +1532,10 @@ static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx, } } } - } while (ep_rv == GRPC_EPOLL_MAX_EVENTS); + + g_current_thread_polling_island = NULL; + gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); + } GPR_ASSERT(pi != NULL); @@ -1868,6 +1995,10 @@ static const grpc_event_engine_vtable vtable = { .kick_poller = kick_poller, + .workqueue_ref = workqueue_ref, + .workqueue_unref = workqueue_unref, + .workqueue_enqueue = workqueue_enqueue, + .shutdown_engine = shutdown_engine, }; @@ -1892,6 +2023,10 @@ const grpc_event_engine_vtable *grpc_init_epoll_linux(void) { return NULL; } + if (!grpc_has_wakeup_fd()) { + return NULL; + } + if (!is_epoll_available()) { return NULL; } diff --git a/src/core/lib/iomgr/ev_poll_and_epoll_posix.c b/src/core/lib/iomgr/ev_poll_and_epoll_posix.c index c2107e5e39..1829440a6e 100644 --- a/src/core/lib/iomgr/ev_poll_and_epoll_posix.c +++ b/src/core/lib/iomgr/ev_poll_and_epoll_posix.c @@ -1989,6 +1989,32 @@ static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, } /******************************************************************************* + * workqueue stubs + */ + +#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG +static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue, + const char *file, int line, + const char *reason) { + return workqueue; +} +static void workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, + const char *file, int line, const char *reason) {} +#else +static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue) { + return workqueue; +} +static void workqueue_unref(grpc_exec_ctx *exec_ctx, + grpc_workqueue *workqueue) {} +#endif + +static void workqueue_enqueue(grpc_exec_ctx *exec_ctx, + grpc_workqueue *workqueue, grpc_closure *closure, + grpc_error *error) { + grpc_exec_ctx_sched(exec_ctx, closure, error, NULL); +} + +/******************************************************************************* * event engine binding */ @@ -2029,6 +2055,10 @@ static const grpc_event_engine_vtable vtable = { .kick_poller = kick_poller, + .workqueue_ref = workqueue_ref, + .workqueue_unref = workqueue_unref, + .workqueue_enqueue = workqueue_enqueue, + .shutdown_engine = shutdown_engine, }; diff --git a/src/core/lib/iomgr/ev_poll_posix.c b/src/core/lib/iomgr/ev_poll_posix.c index 16a5e3083e..27e966c18c 100644 --- a/src/core/lib/iomgr/ev_poll_posix.c +++ b/src/core/lib/iomgr/ev_poll_posix.c @@ -47,10 +47,12 @@ #include <grpc/support/alloc.h> #include <grpc/support/log.h> #include <grpc/support/string_util.h> +#include <grpc/support/thd.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_cv.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" @@ -246,6 +248,28 @@ struct grpc_pollset_set { }; /******************************************************************************* + * condition variable polling definitions + */ + +#define CV_POLL_PERIOD_MS 1000 +#define CV_DEFAULT_TABLE_SIZE 16 + +typedef enum poll_status_t { INPROGRESS, COMPLETED, CANCELLED } poll_status_t; + +typedef struct poll_args { + gpr_refcount refcount; + gpr_cv *cv; + struct pollfd *fds; + nfds_t nfds; + int timeout; + int retval; + int err; + gpr_atm status; +} poll_args; + +cv_fd_table g_cvfds; + +/******************************************************************************* * fd_posix.c */ @@ -961,8 +985,15 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, if (errno != EINTR) { work_combine_error(&error, GRPC_OS_ERROR(errno, "poll")); } + for (i = 2; i < pfd_count; i++) { - fd_end_poll(exec_ctx, &watchers[i], 0, 0, NULL); + if (watchers[i].fd == NULL) { + fd_end_poll(exec_ctx, &watchers[i], 0, 0, NULL); + } else { + // Wake up all the file descriptors, if we have an invalid one + // we can identify it on the next pollset_work() + fd_end_poll(exec_ctx, &watchers[i], 1, 1, pollset); + } } } else if (r == 0) { for (i = 2; i < pfd_count; i++) { @@ -1236,10 +1267,237 @@ static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, } /******************************************************************************* + * workqueue stubs + */ + +#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG +static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue, + const char *file, int line, + const char *reason) { + return workqueue; +} +static void workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, + const char *file, int line, const char *reason) {} +#else +static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue) { + return workqueue; +} +static void workqueue_unref(grpc_exec_ctx *exec_ctx, + grpc_workqueue *workqueue) {} +#endif + +static void workqueue_enqueue(grpc_exec_ctx *exec_ctx, + grpc_workqueue *workqueue, grpc_closure *closure, + grpc_error *error) { + grpc_exec_ctx_sched(exec_ctx, closure, error, NULL); +} + +/******************************************************************************* + * Condition Variable polling extensions + */ + +static void decref_poll_args(poll_args *args) { + if (gpr_unref(&args->refcount)) { + gpr_free(args->fds); + gpr_cv_destroy(args->cv); + gpr_free(args->cv); + gpr_free(args); + } +} + +// Poll in a background thread +static void run_poll(void *arg) { + int timeout, retval; + poll_args *pargs = (poll_args *)arg; + while (gpr_atm_no_barrier_load(&pargs->status) == INPROGRESS) { + if (pargs->timeout < 0) { + timeout = CV_POLL_PERIOD_MS; + } else { + timeout = GPR_MIN(CV_POLL_PERIOD_MS, pargs->timeout); + pargs->timeout -= timeout; + } + retval = g_cvfds.poll(pargs->fds, pargs->nfds, timeout); + if (retval != 0 || pargs->timeout == 0) { + pargs->retval = retval; + pargs->err = errno; + break; + } + } + gpr_mu_lock(&g_cvfds.mu); + if (gpr_atm_no_barrier_load(&pargs->status) == INPROGRESS) { + // Signal main thread that the poll completed + gpr_atm_no_barrier_store(&pargs->status, COMPLETED); + gpr_cv_signal(pargs->cv); + } + decref_poll_args(pargs); + g_cvfds.pollcount--; + if (g_cvfds.shutdown && g_cvfds.pollcount == 0) { + gpr_cv_signal(&g_cvfds.shutdown_complete); + } + gpr_mu_unlock(&g_cvfds.mu); +} + +// This function overrides poll() to handle condition variable wakeup fds +static int cvfd_poll(struct pollfd *fds, nfds_t nfds, int timeout) { + unsigned int i; + int res, idx; + gpr_cv *pollcv; + cv_node *cvn, *prev; + nfds_t nsockfds = 0; + gpr_thd_id t_id; + gpr_thd_options opt; + poll_args *pargs = NULL; + gpr_mu_lock(&g_cvfds.mu); + pollcv = gpr_malloc(sizeof(gpr_cv)); + gpr_cv_init(pollcv); + for (i = 0; i < nfds; i++) { + fds[i].revents = 0; + if (fds[i].fd < 0 && (fds[i].events & POLLIN)) { + idx = FD_TO_IDX(fds[i].fd); + cvn = gpr_malloc(sizeof(cv_node)); + cvn->cv = pollcv; + cvn->next = g_cvfds.cvfds[idx].cvs; + g_cvfds.cvfds[idx].cvs = cvn; + // We should return immediately if there are pending events, + // but we still need to call poll() to check for socket events + if (g_cvfds.cvfds[idx].is_set) { + timeout = 0; + } + } else if (fds[i].fd >= 0) { + nsockfds++; + } + } + + if (nsockfds > 0) { + pargs = gpr_malloc(sizeof(struct poll_args)); + // Both the main thread and calling thread get a reference + gpr_ref_init(&pargs->refcount, 2); + pargs->cv = pollcv; + pargs->fds = gpr_malloc(sizeof(struct pollfd) * nsockfds); + pargs->nfds = nsockfds; + pargs->timeout = timeout; + pargs->retval = 0; + pargs->err = 0; + gpr_atm_no_barrier_store(&pargs->status, INPROGRESS); + idx = 0; + for (i = 0; i < nfds; i++) { + if (fds[i].fd >= 0) { + pargs->fds[idx].fd = fds[i].fd; + pargs->fds[idx].events = fds[i].events; + pargs->fds[idx].revents = 0; + idx++; + } + } + g_cvfds.pollcount++; + opt = gpr_thd_options_default(); + gpr_thd_options_set_detached(&opt); + gpr_thd_new(&t_id, &run_poll, pargs, &opt); + // We want the poll() thread to trigger the deadline, so wait forever here + gpr_cv_wait(pollcv, &g_cvfds.mu, gpr_inf_future(GPR_CLOCK_MONOTONIC)); + if (gpr_atm_no_barrier_load(&pargs->status) == COMPLETED) { + res = pargs->retval; + errno = pargs->err; + } else { + res = 0; + errno = 0; + gpr_atm_no_barrier_store(&pargs->status, CANCELLED); + } + } else { + gpr_timespec deadline = gpr_now(GPR_CLOCK_REALTIME); + deadline = + gpr_time_add(deadline, gpr_time_from_millis(timeout, GPR_TIMESPAN)); + gpr_cv_wait(pollcv, &g_cvfds.mu, deadline); + res = 0; + } + + idx = 0; + for (i = 0; i < nfds; i++) { + if (fds[i].fd < 0 && (fds[i].events & POLLIN)) { + cvn = g_cvfds.cvfds[FD_TO_IDX(fds[i].fd)].cvs; + prev = NULL; + while (cvn->cv != pollcv) { + prev = cvn; + cvn = cvn->next; + GPR_ASSERT(cvn); + } + if (!prev) { + g_cvfds.cvfds[FD_TO_IDX(fds[i].fd)].cvs = cvn->next; + } else { + prev->next = cvn->next; + } + gpr_free(cvn); + + if (g_cvfds.cvfds[FD_TO_IDX(fds[i].fd)].is_set) { + fds[i].revents = POLLIN; + if (res >= 0) res++; + } + } else if (fds[i].fd >= 0 && + gpr_atm_no_barrier_load(&pargs->status) == COMPLETED) { + fds[i].revents = pargs->fds[idx].revents; + idx++; + } + } + + if (pargs) { + decref_poll_args(pargs); + } else { + gpr_cv_destroy(pollcv); + gpr_free(pollcv); + } + gpr_mu_unlock(&g_cvfds.mu); + + return res; +} + +static void global_cv_fd_table_init() { + gpr_mu_init(&g_cvfds.mu); + gpr_mu_lock(&g_cvfds.mu); + gpr_cv_init(&g_cvfds.shutdown_complete); + g_cvfds.shutdown = 0; + g_cvfds.pollcount = 0; + g_cvfds.size = CV_DEFAULT_TABLE_SIZE; + g_cvfds.cvfds = gpr_malloc(sizeof(fd_node) * CV_DEFAULT_TABLE_SIZE); + g_cvfds.free_fds = NULL; + for (int i = 0; i < CV_DEFAULT_TABLE_SIZE; i++) { + g_cvfds.cvfds[i].is_set = 0; + g_cvfds.cvfds[i].cvs = NULL; + g_cvfds.cvfds[i].next_free = g_cvfds.free_fds; + g_cvfds.free_fds = &g_cvfds.cvfds[i]; + } + // Override the poll function with one that supports cvfds + g_cvfds.poll = grpc_poll_function; + grpc_poll_function = &cvfd_poll; + gpr_mu_unlock(&g_cvfds.mu); +} + +static void global_cv_fd_table_shutdown() { + gpr_mu_lock(&g_cvfds.mu); + g_cvfds.shutdown = 1; + // Attempt to wait for all abandoned poll() threads to terminate + // Not doing so will result in reported memory leaks + if (g_cvfds.pollcount > 0) { + int res = gpr_cv_wait(&g_cvfds.shutdown_complete, &g_cvfds.mu, + gpr_time_add(gpr_now(GPR_CLOCK_REALTIME), + gpr_time_from_seconds(3, GPR_TIMESPAN))); + GPR_ASSERT(res == 0); + } + gpr_cv_destroy(&g_cvfds.shutdown_complete); + grpc_poll_function = g_cvfds.poll; + gpr_free(g_cvfds.cvfds); + gpr_mu_unlock(&g_cvfds.mu); + gpr_mu_destroy(&g_cvfds.mu); +} + +/******************************************************************************* * event engine binding */ -static void shutdown_engine(void) { pollset_global_shutdown(); } +static void shutdown_engine(void) { + pollset_global_shutdown(); + if (grpc_cv_wakeup_fds_enabled()) { + global_cv_fd_table_shutdown(); + } +} static const grpc_event_engine_vtable vtable = { .pollset_size = sizeof(grpc_pollset), @@ -1273,11 +1531,29 @@ static const grpc_event_engine_vtable vtable = { .kick_poller = kick_poller, + .workqueue_ref = workqueue_ref, + .workqueue_unref = workqueue_unref, + .workqueue_enqueue = workqueue_enqueue, + .shutdown_engine = shutdown_engine, }; const grpc_event_engine_vtable *grpc_init_poll_posix(void) { + if (!grpc_has_wakeup_fd()) { + return NULL; + } + if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) { + return NULL; + } + return &vtable; +} + +const grpc_event_engine_vtable *grpc_init_poll_cv_posix(void) { + global_cv_fd_table_init(); + grpc_enable_cv_wakeup_fds(1); if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) { + global_cv_fd_table_shutdown(); + grpc_enable_cv_wakeup_fds(0); return NULL; } return &vtable; diff --git a/src/core/lib/iomgr/ev_poll_posix.h b/src/core/lib/iomgr/ev_poll_posix.h index 291736a2db..202ffca14c 100644 --- a/src/core/lib/iomgr/ev_poll_posix.h +++ b/src/core/lib/iomgr/ev_poll_posix.h @@ -37,5 +37,6 @@ #include "src/core/lib/iomgr/ev_posix.h" const grpc_event_engine_vtable *grpc_init_poll_posix(void); +const grpc_event_engine_vtable *grpc_init_poll_cv_posix(void); #endif /* GRPC_CORE_LIB_IOMGR_EV_POLL_POSIX_H */ diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c index 6536672685..9857b0bce9 100644 --- a/src/core/lib/iomgr/ev_posix.c +++ b/src/core/lib/iomgr/ev_posix.c @@ -66,6 +66,7 @@ typedef struct { static const event_engine_factory g_factories[] = { {"epoll", grpc_init_epoll_linux}, {"poll", grpc_init_poll_posix}, + {"poll-cv", grpc_init_poll_cv_posix}, {"legacy", grpc_init_poll_and_epoll_posix}, }; @@ -258,4 +259,27 @@ void grpc_pollset_set_del_fd(grpc_exec_ctx *exec_ctx, grpc_error *grpc_kick_poller(void) { return g_event_engine->kick_poller(); } +#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG +grpc_workqueue *grpc_workqueue_ref(grpc_workqueue *workqueue, const char *file, + int line, const char *reason) { + return g_event_engine->workqueue_ref(workqueue, file, line, reason); +} +void grpc_workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, + const char *file, int line, const char *reason) { + g_event_engine->workqueue_unref(exec_ctx, workqueue, file, line, reason); +} +#else +grpc_workqueue *grpc_workqueue_ref(grpc_workqueue *workqueue) { + return g_event_engine->workqueue_ref(workqueue); +} +void grpc_workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue) { + g_event_engine->workqueue_unref(exec_ctx, workqueue); +} +#endif + +void grpc_workqueue_enqueue(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, + grpc_closure *closure, grpc_error *error) { + g_event_engine->workqueue_enqueue(exec_ctx, workqueue, closure, error); +} + #endif // GPR_POSIX_SOCKET diff --git a/src/core/lib/iomgr/ev_posix.h b/src/core/lib/iomgr/ev_posix.h index c2aa1756ea..2fdef06838 100644 --- a/src/core/lib/iomgr/ev_posix.h +++ b/src/core/lib/iomgr/ev_posix.h @@ -40,6 +40,7 @@ #include "src/core/lib/iomgr/pollset.h" #include "src/core/lib/iomgr/pollset_set.h" #include "src/core/lib/iomgr/wakeup_fd_posix.h" +#include "src/core/lib/iomgr/workqueue.h" typedef struct grpc_fd grpc_fd; @@ -95,6 +96,18 @@ typedef struct grpc_event_engine_vtable { grpc_error *(*kick_poller)(void); void (*shutdown_engine)(void); + +#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG + grpc_workqueue *(*workqueue_ref)(grpc_workqueue *workqueue, const char *file, + int line, const char *reason); + void (*workqueue_unref)(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, + const char *file, int line, const char *reason); +#else + grpc_workqueue *(*workqueue_ref)(grpc_workqueue *workqueue); + void (*workqueue_unref)(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue); +#endif + void (*workqueue_enqueue)(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, + grpc_closure *closure, grpc_error *error); } grpc_event_engine_vtable; void grpc_event_engine_init(void); diff --git a/src/core/lib/iomgr/exec_ctx.c b/src/core/lib/iomgr/exec_ctx.c index ac7785ec13..604713e578 100644 --- a/src/core/lib/iomgr/exec_ctx.c +++ b/src/core/lib/iomgr/exec_ctx.c @@ -37,6 +37,7 @@ #include <grpc/support/sync.h> #include <grpc/support/thd.h> +#include "src/core/lib/iomgr/combiner.h" #include "src/core/lib/iomgr/workqueue.h" #include "src/core/lib/profiling/timers.h" @@ -60,18 +61,43 @@ bool grpc_always_ready_to_finish(grpc_exec_ctx *exec_ctx, void *arg_ignored) { bool grpc_exec_ctx_flush(grpc_exec_ctx *exec_ctx) { bool did_something = 0; GPR_TIMER_BEGIN("grpc_exec_ctx_flush", 0); - while (!grpc_closure_list_empty(exec_ctx->closure_list)) { - grpc_closure *c = exec_ctx->closure_list.head; - exec_ctx->closure_list.head = exec_ctx->closure_list.tail = NULL; - while (c != NULL) { - grpc_closure *next = c->next_data.next; - grpc_error *error = c->error; - did_something = true; - GPR_TIMER_BEGIN("grpc_exec_ctx_flush.cb", 0); + for (;;) { + if (!grpc_closure_list_empty(exec_ctx->closure_list)) { + grpc_closure *c = exec_ctx->closure_list.head; + exec_ctx->closure_list.head = exec_ctx->closure_list.tail = NULL; + while (c != NULL) { + grpc_closure *next = c->next_data.next; + did_something = true; + grpc_closure_run(exec_ctx, c, c->error_data.error); + c = next; + } + } else if (!grpc_combiner_continue_exec_ctx(exec_ctx)) { + break; + } + } + GPR_ASSERT(exec_ctx->active_combiner == NULL); + if (exec_ctx->stealing_from_workqueue != NULL) { + if (grpc_exec_ctx_ready_to_finish(exec_ctx)) { + grpc_workqueue_enqueue(exec_ctx, exec_ctx->stealing_from_workqueue, + exec_ctx->stolen_closure, + exec_ctx->stolen_closure->error_data.error); + GRPC_WORKQUEUE_UNREF(exec_ctx, exec_ctx->stealing_from_workqueue, + "exec_ctx_sched"); + exec_ctx->stealing_from_workqueue = NULL; + exec_ctx->stolen_closure = NULL; + } else { + grpc_closure *c = exec_ctx->stolen_closure; + GRPC_WORKQUEUE_UNREF(exec_ctx, exec_ctx->stealing_from_workqueue, + "exec_ctx_sched"); + exec_ctx->stealing_from_workqueue = NULL; + exec_ctx->stolen_closure = NULL; + grpc_error *error = c->error_data.error; + GPR_TIMER_BEGIN("grpc_exec_ctx_flush.stolen_cb", 0); c->cb(exec_ctx, c->cb_arg, error); GRPC_ERROR_UNREF(error); - GPR_TIMER_END("grpc_exec_ctx_flush.cb", 0); - c = next; + GPR_TIMER_END("grpc_exec_ctx_flush.stolen_cb", 0); + grpc_exec_ctx_flush(exec_ctx); + did_something = true; } } GPR_TIMER_END("grpc_exec_ctx_flush", 0); @@ -86,12 +112,25 @@ void grpc_exec_ctx_finish(grpc_exec_ctx *exec_ctx) { void grpc_exec_ctx_sched(grpc_exec_ctx *exec_ctx, grpc_closure *closure, grpc_error *error, grpc_workqueue *offload_target_or_null) { + GPR_TIMER_BEGIN("grpc_exec_ctx_sched", 0); if (offload_target_or_null == NULL) { grpc_closure_list_append(&exec_ctx->closure_list, closure, error); - } else { + } else if (exec_ctx->stealing_from_workqueue == NULL) { + exec_ctx->stealing_from_workqueue = offload_target_or_null; + closure->error_data.error = error; + exec_ctx->stolen_closure = closure; + } else if (exec_ctx->stealing_from_workqueue != offload_target_or_null) { grpc_workqueue_enqueue(exec_ctx, offload_target_or_null, closure, error); GRPC_WORKQUEUE_UNREF(exec_ctx, offload_target_or_null, "exec_ctx_sched"); + } else { /* stealing_from_workqueue == offload_target_or_null */ + grpc_workqueue_enqueue(exec_ctx, offload_target_or_null, + exec_ctx->stolen_closure, + exec_ctx->stolen_closure->error_data.error); + closure->error_data.error = error; + exec_ctx->stolen_closure = closure; + GRPC_WORKQUEUE_UNREF(exec_ctx, offload_target_or_null, "exec_ctx_sched"); } + GPR_TIMER_END("grpc_exec_ctx_sched", 0); } void grpc_exec_ctx_enqueue_list(grpc_exec_ctx *exec_ctx, diff --git a/src/core/lib/iomgr/exec_ctx.h b/src/core/lib/iomgr/exec_ctx.h index 4d20ecf922..7e50cb9825 100644 --- a/src/core/lib/iomgr/exec_ctx.h +++ b/src/core/lib/iomgr/exec_ctx.h @@ -66,15 +66,33 @@ typedef struct grpc_combiner grpc_combiner; #ifndef GRPC_EXECUTION_CONTEXT_SANITIZER struct grpc_exec_ctx { grpc_closure_list closure_list; + /** The workqueue we're stealing work from. + As items are queued to the execution context, we try to steal one + workqueue item and execute it inline (assuming the exec_ctx is not + finished) - doing so does not invalidate the workqueue's contract, and + provides a small latency win in cases where we get a hit */ + grpc_workqueue *stealing_from_workqueue; + /** The workqueue item that was stolen from the workqueue above. When new + items are scheduled to be offloaded to that workqueue, we need to update + this like a 1-deep fifo to maintain the invariant that workqueue items + queued by one thread are started in order */ + grpc_closure *stolen_closure; /** currently active combiner: updated only via combiner.c */ grpc_combiner *active_combiner; + /** last active combiner in the active combiner list */ + grpc_combiner *last_combiner; bool cached_ready_to_finish; void *check_ready_to_finish_arg; bool (*check_ready_to_finish)(grpc_exec_ctx *exec_ctx, void *arg); }; +/* initializer for grpc_exec_ctx: + prefer to use GRPC_EXEC_CTX_INIT whenever possible */ #define GRPC_EXEC_CTX_INIT_WITH_FINISH_CHECK(finish_check, finish_check_arg) \ - { GRPC_CLOSURE_LIST_INIT, NULL, false, finish_check_arg, finish_check } + { \ + GRPC_CLOSURE_LIST_INIT, NULL, NULL, NULL, NULL, false, finish_check_arg, \ + finish_check \ + } #else struct grpc_exec_ctx { bool cached_ready_to_finish; @@ -85,8 +103,10 @@ struct grpc_exec_ctx { { false, finish_check_arg, finish_check } #endif +/* initialize an execution context at the top level of an API call into grpc + (this is safe to use elsewhere, though possibly not as efficient) */ #define GRPC_EXEC_CTX_INIT \ - GRPC_EXEC_CTX_INIT_WITH_FINISH_CHECK(grpc_never_ready_to_finish, NULL) + GRPC_EXEC_CTX_INIT_WITH_FINISH_CHECK(grpc_always_ready_to_finish, NULL) /** Flush any work that has been enqueued onto this grpc_exec_ctx. * Caller must guarantee that no interfering locks are held. diff --git a/src/core/lib/iomgr/iomgr.c b/src/core/lib/iomgr/iomgr.c index d67d388b8c..4fd83e0b22 100644 --- a/src/core/lib/iomgr/iomgr.c +++ b/src/core/lib/iomgr/iomgr.c @@ -112,6 +112,14 @@ void grpc_iomgr_shutdown(void) { continue; } if (g_root_object.next != &g_root_object) { + if (grpc_iomgr_abort_on_leaks()) { + gpr_log(GPR_DEBUG, "Failed to free %" PRIuPTR + " iomgr objects before shutdown deadline: " + "memory leaks are likely", + count_objects()); + dump_objects("LEAKED"); + abort(); + } gpr_timespec short_deadline = gpr_time_add( gpr_now(GPR_CLOCK_REALTIME), gpr_time_from_millis(100, GPR_TIMESPAN)); if (gpr_cv_wait(&g_rcv, &g_mu, short_deadline)) { @@ -122,9 +130,6 @@ void grpc_iomgr_shutdown(void) { "memory leaks are likely", count_objects()); dump_objects("LEAKED"); - if (grpc_iomgr_abort_on_leaks()) { - abort(); - } } break; } diff --git a/src/core/lib/iomgr/tcp_posix.c b/src/core/lib/iomgr/tcp_posix.c index 92767721d5..00fd77679a 100644 --- a/src/core/lib/iomgr/tcp_posix.c +++ b/src/core/lib/iomgr/tcp_posix.c @@ -177,7 +177,7 @@ static void call_read_cb(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp, tcp->read_cb = NULL; tcp->incoming_buffer = NULL; - grpc_exec_ctx_sched(exec_ctx, cb, error, NULL); + grpc_closure_run(exec_ctx, cb, error); } #define MAX_READ_IOVEC 4 @@ -209,11 +209,11 @@ static void tcp_continue_read(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) { msg.msg_controllen = 0; msg.msg_flags = 0; - GPR_TIMER_BEGIN("recvmsg", 1); + GPR_TIMER_BEGIN("recvmsg", 0); do { read_bytes = recvmsg(tcp->fd, &msg, 0); } while (read_bytes < 0 && errno == EINTR); - GPR_TIMER_END("recvmsg", 0); + GPR_TIMER_END("recvmsg", read_bytes >= 0); if (read_bytes < 0) { /* NB: After calling call_read_cb a parallel call of the read handler may @@ -392,11 +392,8 @@ static void tcp_handle_write(grpc_exec_ctx *exec_ctx, void *arg /* grpc_tcp */, grpc_error_free_string(str); } - GPR_TIMER_BEGIN("tcp_handle_write.cb", 0); - cb->cb(exec_ctx, cb->cb_arg, error); - GPR_TIMER_END("tcp_handle_write.cb", 0); + grpc_closure_run(exec_ctx, cb, error); TCP_UNREF(exec_ctx, tcp, "write"); - GRPC_ERROR_UNREF(error); } } diff --git a/src/core/lib/iomgr/timer.h b/src/core/lib/iomgr/timer.h index a825d2a28b..5a9a177963 100644 --- a/src/core/lib/iomgr/timer.h +++ b/src/core/lib/iomgr/timer.h @@ -49,11 +49,11 @@ typedef struct grpc_timer { } grpc_timer; /* Initialize *timer. When expired or canceled, timer_cb will be called with - *timer_cb_arg and status to indicate if it expired (SUCCESS) or was - canceled (CANCELLED). timer_cb is guaranteed to be called exactly once, - and application code should check the status to determine how it was - invoked. The application callback is also responsible for maintaining - information about when to free up any user-level state. */ + *timer_cb_arg and error set to indicate if it expired (GRPC_ERROR_NONE) or + was canceled (GRPC_ERROR_CANCELLED). timer_cb is guaranteed to be called + exactly once, and application code should check the error to determine + how it was invoked. The application callback is also responsible for + maintaining information about when to free up any user-level state. */ void grpc_timer_init(grpc_exec_ctx *exec_ctx, grpc_timer *timer, gpr_timespec deadline, grpc_iomgr_cb_func timer_cb, void *timer_cb_arg, gpr_timespec now); @@ -74,8 +74,8 @@ void grpc_timer_init(grpc_exec_ctx *exec_ctx, grpc_timer *timer, In all of these cases, the cancellation is still considered successful. They are essentially distinguished in that the timer_cb will be run - exactly once from either the cancellation (with status CANCELLED) - or from the activation (with status SUCCESS) + exactly once from either the cancellation (with error GRPC_ERROR_CANCELLED) + or from the activation (with error GRPC_ERROR_NONE). Note carefully that the callback function MAY occur in the same callstack as grpc_timer_cancel. It's expected that most timers will be cancelled (their @@ -83,14 +83,13 @@ void grpc_timer_init(grpc_exec_ctx *exec_ctx, grpc_timer *timer, that cancellation costs as little as possible. Making callbacks run inline matches this aim. - Requires: cancel() must happen after add() on a given timer */ + Requires: cancel() must happen after init() on a given timer */ void grpc_timer_cancel(grpc_exec_ctx *exec_ctx, grpc_timer *timer); /* iomgr internal api for dealing with timers */ /* Check for timers to be run, and run them. Return true if timer callbacks were executed. - Drops drop_mu if it is non-null before executing callbacks. If next is non-null, TRY to update *next with the next running timer IF that timer occurs before *next current value. *next is never guaranteed to be updated on any given execution; however, diff --git a/src/core/lib/iomgr/wakeup_fd_cv.c b/src/core/lib/iomgr/wakeup_fd_cv.c new file mode 100644 index 0000000000..b4165208ed --- /dev/null +++ b/src/core/lib/iomgr/wakeup_fd_cv.c @@ -0,0 +1,118 @@ +/* + * + * Copyright 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_WAKEUP_FD + +#include "src/core/lib/iomgr/wakeup_fd_cv.h" + +#include <errno.h> +#include <string.h> + +#include <grpc/support/alloc.h> +#include <grpc/support/log.h> +#include <grpc/support/sync.h> +#include <grpc/support/thd.h> +#include <grpc/support/time.h> +#include <grpc/support/useful.h> + +#define MAX_TABLE_RESIZE 256 + +extern cv_fd_table g_cvfds; + +static grpc_error* cv_fd_init(grpc_wakeup_fd* fd_info) { + unsigned int i, newsize; + int idx; + gpr_mu_lock(&g_cvfds.mu); + if (!g_cvfds.free_fds) { + newsize = GPR_MIN(g_cvfds.size * 2, g_cvfds.size + MAX_TABLE_RESIZE); + g_cvfds.cvfds = gpr_realloc(g_cvfds.cvfds, sizeof(fd_node) * newsize); + for (i = g_cvfds.size; i < newsize; i++) { + g_cvfds.cvfds[i].is_set = 0; + g_cvfds.cvfds[i].cvs = NULL; + g_cvfds.cvfds[i].next_free = g_cvfds.free_fds; + g_cvfds.free_fds = &g_cvfds.cvfds[i]; + } + g_cvfds.size = newsize; + } + + idx = (int)(g_cvfds.free_fds - g_cvfds.cvfds); + g_cvfds.free_fds = g_cvfds.free_fds->next_free; + g_cvfds.cvfds[idx].cvs = NULL; + g_cvfds.cvfds[idx].is_set = 0; + fd_info->read_fd = IDX_TO_FD(idx); + fd_info->write_fd = -1; + gpr_mu_unlock(&g_cvfds.mu); + return GRPC_ERROR_NONE; +} + +static grpc_error* cv_fd_wakeup(grpc_wakeup_fd* fd_info) { + cv_node* cvn; + gpr_mu_lock(&g_cvfds.mu); + g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)].is_set = 1; + cvn = g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)].cvs; + while (cvn) { + gpr_cv_signal(cvn->cv); + cvn = cvn->next; + } + gpr_mu_unlock(&g_cvfds.mu); + return GRPC_ERROR_NONE; +} + +static grpc_error* cv_fd_consume(grpc_wakeup_fd* fd_info) { + gpr_mu_lock(&g_cvfds.mu); + g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)].is_set = 0; + gpr_mu_unlock(&g_cvfds.mu); + return GRPC_ERROR_NONE; +} + +static void cv_fd_destroy(grpc_wakeup_fd* fd_info) { + if (fd_info->read_fd == 0) { + return; + } + gpr_mu_lock(&g_cvfds.mu); + // Assert that there are no active pollers + GPR_ASSERT(!g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)].cvs); + g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)].next_free = g_cvfds.free_fds; + g_cvfds.free_fds = &g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)]; + gpr_mu_unlock(&g_cvfds.mu); +} + +static int cv_check_availability(void) { return 1; } + +const grpc_wakeup_fd_vtable grpc_cv_wakeup_fd_vtable = { + cv_fd_init, cv_fd_consume, cv_fd_wakeup, cv_fd_destroy, + cv_check_availability}; + +#endif /* GPR_POSIX_WAKUP_FD */ diff --git a/src/core/lib/iomgr/workqueue_posix.h b/src/core/lib/iomgr/wakeup_fd_cv.h index 03ee21cef7..ac16be1750 100644 --- a/src/core/lib/iomgr/workqueue_posix.h +++ b/src/core/lib/iomgr/wakeup_fd_cv.h @@ -1,6 +1,6 @@ /* * - * Copyright 2015, Google Inc. + * Copyright 2016, Google Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without @@ -31,31 +31,50 @@ * */ -#ifndef GRPC_CORE_LIB_IOMGR_WORKQUEUE_POSIX_H -#define GRPC_CORE_LIB_IOMGR_WORKQUEUE_POSIX_H +/* + * wakeup_fd_cv uses condition variables to implement wakeup fds. + * + * It is intended for use only in cases when eventfd() and pipe() are not + * available. It can only be used with the "poll" engine. + * + * Implementation: + * A global table of cv wakeup fds is mantained. A cv wakeup fd is a negative + * file descriptor. poll() is then run in a background thread with only the + * real socket fds while we wait on a condition variable trigged by either the + * poll() completion or a wakeup_fd() call. + * + */ + +#ifndef GRPC_CORE_LIB_IOMGR_WAKEUP_FD_CV_H +#define GRPC_CORE_LIB_IOMGR_WAKEUP_FD_CV_H -#include "src/core/lib/iomgr/wakeup_fd_posix.h" -#include "src/core/lib/support/mpscq.h" +#include <grpc/support/sync.h> -struct grpc_fd; +#include "src/core/lib/iomgr/ev_posix.h" -struct grpc_workqueue { - gpr_refcount refs; - gpr_mpscq queue; - // state is: - // lower bit - zero if orphaned - // other bits - number of items enqueued - gpr_atm state; +#define FD_TO_IDX(fd) (-(fd)-1) +#define IDX_TO_FD(idx) (-(idx)-1) - grpc_wakeup_fd wakeup_fd; - struct grpc_fd *wakeup_read_fd; +typedef struct cv_node { + gpr_cv* cv; + struct cv_node* next; +} cv_node; - grpc_closure read_closure; -}; +typedef struct fd_node { + int is_set; + cv_node* cvs; + struct fd_node* next_free; +} fd_node; -/** Create a work queue. Returns an error if creation fails. If creation - succeeds, sets *workqueue to point to it. */ -grpc_error *grpc_workqueue_create(grpc_exec_ctx *exec_ctx, - grpc_workqueue **workqueue); +typedef struct cv_fd_table { + gpr_mu mu; + int pollcount; + int shutdown; + gpr_cv shutdown_complete; + fd_node* cvfds; + fd_node* free_fds; + unsigned int size; + grpc_poll_function_type poll; +} cv_fd_table; -#endif /* GRPC_CORE_LIB_IOMGR_WORKQUEUE_POSIX_H */ +#endif /* GRPC_CORE_LIB_IOMGR_WAKEUP_FD_CV_H */ diff --git a/src/core/lib/iomgr/wakeup_fd_pipe.c b/src/core/lib/iomgr/wakeup_fd_pipe.c index 4e5dbdcb73..d0ea216aa0 100644 --- a/src/core/lib/iomgr/wakeup_fd_pipe.c +++ b/src/core/lib/iomgr/wakeup_fd_pipe.c @@ -47,11 +47,10 @@ static grpc_error* pipe_init(grpc_wakeup_fd* fd_info) { int pipefd[2]; - /* TODO(klempner): Make this nonfatal */ int r = pipe(pipefd); if (0 != r) { gpr_log(GPR_ERROR, "pipe creation failed (%d): %s", errno, strerror(errno)); - abort(); + return GRPC_OS_ERROR(errno, "pipe"); } grpc_error* err; err = grpc_set_socket_nonblocking(pipefd[0], 1); @@ -95,8 +94,13 @@ static void pipe_destroy(grpc_wakeup_fd* fd_info) { } static int pipe_check_availability(void) { - /* Assume that pipes are always available. */ - return 1; + grpc_wakeup_fd fd; + if (pipe_init(&fd) == GRPC_ERROR_NONE) { + pipe_destroy(&fd); + return 1; + } else { + return 0; + } } const grpc_wakeup_fd_vtable grpc_pipe_wakeup_fd_vtable = { diff --git a/src/core/lib/iomgr/wakeup_fd_posix.c b/src/core/lib/iomgr/wakeup_fd_posix.c index 046208abc8..5c894bef37 100644 --- a/src/core/lib/iomgr/wakeup_fd_posix.c +++ b/src/core/lib/iomgr/wakeup_fd_posix.c @@ -36,37 +36,66 @@ #ifdef GPR_POSIX_WAKEUP_FD #include <stddef.h> +#include "src/core/lib/iomgr/wakeup_fd_cv.h" #include "src/core/lib/iomgr/wakeup_fd_pipe.h" #include "src/core/lib/iomgr/wakeup_fd_posix.h" +extern grpc_wakeup_fd_vtable grpc_cv_wakeup_fd_vtable; static const grpc_wakeup_fd_vtable *wakeup_fd_vtable = NULL; + int grpc_allow_specialized_wakeup_fd = 1; +int grpc_allow_pipe_wakeup_fd = 1; + +int has_real_wakeup_fd = 1; +int cv_wakeup_fds_enabled = 0; void grpc_wakeup_fd_global_init(void) { if (grpc_allow_specialized_wakeup_fd && grpc_specialized_wakeup_fd_vtable.check_availability()) { wakeup_fd_vtable = &grpc_specialized_wakeup_fd_vtable; - } else { + } else if (grpc_allow_pipe_wakeup_fd && + grpc_pipe_wakeup_fd_vtable.check_availability()) { wakeup_fd_vtable = &grpc_pipe_wakeup_fd_vtable; + } else { + has_real_wakeup_fd = 0; } } void grpc_wakeup_fd_global_destroy(void) { wakeup_fd_vtable = NULL; } +int grpc_has_wakeup_fd(void) { return has_real_wakeup_fd; } + +int grpc_cv_wakeup_fds_enabled(void) { return cv_wakeup_fds_enabled; } + +void grpc_enable_cv_wakeup_fds(int enable) { cv_wakeup_fds_enabled = enable; } + grpc_error *grpc_wakeup_fd_init(grpc_wakeup_fd *fd_info) { + if (cv_wakeup_fds_enabled) { + return grpc_cv_wakeup_fd_vtable.init(fd_info); + } return wakeup_fd_vtable->init(fd_info); } grpc_error *grpc_wakeup_fd_consume_wakeup(grpc_wakeup_fd *fd_info) { + if (cv_wakeup_fds_enabled) { + return grpc_cv_wakeup_fd_vtable.consume(fd_info); + } return wakeup_fd_vtable->consume(fd_info); } grpc_error *grpc_wakeup_fd_wakeup(grpc_wakeup_fd *fd_info) { + if (cv_wakeup_fds_enabled) { + return grpc_cv_wakeup_fd_vtable.wakeup(fd_info); + } return wakeup_fd_vtable->wakeup(fd_info); } void grpc_wakeup_fd_destroy(grpc_wakeup_fd *fd_info) { - wakeup_fd_vtable->destroy(fd_info); + if (cv_wakeup_fds_enabled) { + grpc_cv_wakeup_fd_vtable.destroy(fd_info); + } else { + wakeup_fd_vtable->destroy(fd_info); + } } #endif /* GPR_POSIX_WAKEUP_FD */ diff --git a/src/core/lib/iomgr/wakeup_fd_posix.h b/src/core/lib/iomgr/wakeup_fd_posix.h index e269f242d8..71d32d97ba 100644 --- a/src/core/lib/iomgr/wakeup_fd_posix.h +++ b/src/core/lib/iomgr/wakeup_fd_posix.h @@ -71,6 +71,10 @@ void grpc_wakeup_fd_global_destroy(void); * purposes only.*/ void grpc_wakeup_fd_global_init_force_fallback(void); +int grpc_has_wakeup_fd(void); +int grpc_cv_wakeup_fds_enabled(void); +void grpc_enable_cv_wakeup_fds(int enable); + typedef struct grpc_wakeup_fd grpc_wakeup_fd; typedef struct grpc_wakeup_fd_vtable { @@ -88,6 +92,7 @@ struct grpc_wakeup_fd { }; extern int grpc_allow_specialized_wakeup_fd; +extern int grpc_allow_pipe_wakeup_fd; #define GRPC_WAKEUP_FD_GET_READ_FD(fd_info) ((fd_info)->read_fd) diff --git a/src/core/lib/iomgr/workqueue.h b/src/core/lib/iomgr/workqueue.h index b2805dc66c..5b96d1d851 100644 --- a/src/core/lib/iomgr/workqueue.h +++ b/src/core/lib/iomgr/workqueue.h @@ -40,10 +40,6 @@ #include "src/core/lib/iomgr/pollset.h" #include "src/core/lib/iomgr/pollset_set.h" -#ifdef GPR_POSIX_SOCKET -#include "src/core/lib/iomgr/workqueue_posix.h" -#endif - #ifdef GPR_WINDOWS #include "src/core/lib/iomgr/workqueue_windows.h" #endif @@ -58,20 +54,20 @@ string will be printed alongside the refcount. When it is not defined, the string will be discarded at compilation time. */ -//#define GRPC_WORKQUEUE_REFCOUNT_DEBUG +/*#define GRPC_WORKQUEUE_REFCOUNT_DEBUG*/ #ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG #define GRPC_WORKQUEUE_REF(p, r) \ - (grpc_workqueue_ref((p), __FILE__, __LINE__, (r)), (p)) + grpc_workqueue_ref((p), __FILE__, __LINE__, (r)) #define GRPC_WORKQUEUE_UNREF(exec_ctx, p, r) \ grpc_workqueue_unref((exec_ctx), (p), __FILE__, __LINE__, (r)) -void grpc_workqueue_ref(grpc_workqueue *workqueue, const char *file, int line, - const char *reason); +grpc_workqueue *grpc_workqueue_ref(grpc_workqueue *workqueue, const char *file, + int line, const char *reason); void grpc_workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, const char *file, int line, const char *reason); #else -#define GRPC_WORKQUEUE_REF(p, r) (grpc_workqueue_ref((p)), (p)) +#define GRPC_WORKQUEUE_REF(p, r) grpc_workqueue_ref((p)) #define GRPC_WORKQUEUE_UNREF(cl, p, r) grpc_workqueue_unref((cl), (p)) -void grpc_workqueue_ref(grpc_workqueue *workqueue); +grpc_workqueue *grpc_workqueue_ref(grpc_workqueue *workqueue); void grpc_workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue); #endif diff --git a/src/core/lib/iomgr/workqueue_posix.c b/src/core/lib/iomgr/workqueue_posix.c deleted file mode 100644 index ecfea68f56..0000000000 --- a/src/core/lib/iomgr/workqueue_posix.c +++ /dev/null @@ -1,196 +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/workqueue.h" - -#include <stdio.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" - -static void on_readable(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error); - -grpc_error *grpc_workqueue_create(grpc_exec_ctx *exec_ctx, - grpc_workqueue **workqueue) { - char name[32]; - *workqueue = gpr_malloc(sizeof(grpc_workqueue)); - gpr_ref_init(&(*workqueue)->refs, 1); - gpr_atm_no_barrier_store(&(*workqueue)->state, 1); - grpc_error *err = grpc_wakeup_fd_init(&(*workqueue)->wakeup_fd); - if (err != GRPC_ERROR_NONE) { - gpr_free(*workqueue); - return err; - } - sprintf(name, "workqueue:%p", (void *)(*workqueue)); - (*workqueue)->wakeup_read_fd = grpc_fd_create( - GRPC_WAKEUP_FD_GET_READ_FD(&(*workqueue)->wakeup_fd), name); - gpr_mpscq_init(&(*workqueue)->queue); - grpc_closure_init(&(*workqueue)->read_closure, on_readable, *workqueue); - grpc_fd_notify_on_read(exec_ctx, (*workqueue)->wakeup_read_fd, - &(*workqueue)->read_closure); - return GRPC_ERROR_NONE; -} - -static void workqueue_destroy(grpc_exec_ctx *exec_ctx, - grpc_workqueue *workqueue) { - grpc_fd_shutdown(exec_ctx, workqueue->wakeup_read_fd); -} - -static void workqueue_orphan(grpc_exec_ctx *exec_ctx, - grpc_workqueue *workqueue) { - if (gpr_atm_full_fetch_add(&workqueue->state, -1) == 1) { - workqueue_destroy(exec_ctx, workqueue); - } -} - -#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG -void grpc_workqueue_ref(grpc_workqueue *workqueue, const char *file, int line, - const char *reason) { - if (workqueue == NULL) return; - gpr_log(file, line, GPR_LOG_SEVERITY_DEBUG, "WORKQUEUE:%p ref %d -> %d %s", - workqueue, (int)workqueue->refs.count, (int)workqueue->refs.count + 1, - reason); - gpr_ref(&workqueue->refs); -} -#else -void grpc_workqueue_ref(grpc_workqueue *workqueue) { - if (workqueue == NULL) return; - gpr_ref(&workqueue->refs); -} -#endif - -#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG -void grpc_workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, - const char *file, int line, const char *reason) { - if (workqueue == NULL) return; - gpr_log(file, line, GPR_LOG_SEVERITY_DEBUG, "WORKQUEUE:%p unref %d -> %d %s", - workqueue, (int)workqueue->refs.count, (int)workqueue->refs.count - 1, - reason); - if (gpr_unref(&workqueue->refs)) { - workqueue_orphan(exec_ctx, workqueue); - } -} -#else -void grpc_workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue) { - if (workqueue == NULL) return; - if (gpr_unref(&workqueue->refs)) { - workqueue_orphan(exec_ctx, workqueue); - } -} -#endif - -static void drain(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue) { - abort(); -} - -static void wakeup(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue) { - GPR_TIMER_MARK("workqueue.wakeup", 0); - grpc_error *err = grpc_wakeup_fd_wakeup(&workqueue->wakeup_fd); - if (!GRPC_LOG_IF_ERROR("wakeupfd_wakeup", err)) { - drain(exec_ctx, workqueue); - } -} - -static void on_readable(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { - GPR_TIMER_BEGIN("workqueue.on_readable", 0); - - grpc_workqueue *workqueue = arg; - - if (error != GRPC_ERROR_NONE) { - /* HACK: let wakeup_fd code know that we stole the fd */ - workqueue->wakeup_fd.read_fd = 0; - grpc_wakeup_fd_destroy(&workqueue->wakeup_fd); - grpc_fd_orphan(exec_ctx, workqueue->wakeup_read_fd, NULL, NULL, "destroy"); - GPR_ASSERT(gpr_atm_no_barrier_load(&workqueue->state) == 0); - gpr_free(workqueue); - } else { - error = grpc_wakeup_fd_consume_wakeup(&workqueue->wakeup_fd); - gpr_mpscq_node *n = gpr_mpscq_pop(&workqueue->queue); - if (error == GRPC_ERROR_NONE) { - grpc_fd_notify_on_read(exec_ctx, workqueue->wakeup_read_fd, - &workqueue->read_closure); - } else { - /* recurse to get error handling */ - on_readable(exec_ctx, arg, error); - } - if (n == NULL) { - /* try again - queue in an inconsistant state */ - wakeup(exec_ctx, workqueue); - } else { - switch (gpr_atm_full_fetch_add(&workqueue->state, -2)) { - case 3: // had one count, one unorphaned --> done, unorphaned - break; - case 2: // had one count, one orphaned --> done, orphaned - workqueue_destroy(exec_ctx, workqueue); - break; - case 1: - case 0: - // these values are illegal - representing an already done or - // deleted workqueue - GPR_UNREACHABLE_CODE(break); - default: - // schedule a wakeup since there's more to do - wakeup(exec_ctx, workqueue); - } - grpc_closure *cl = (grpc_closure *)n; - grpc_error *clerr = cl->error; - cl->cb(exec_ctx, cl->cb_arg, clerr); - GRPC_ERROR_UNREF(clerr); - } - } - - GPR_TIMER_END("workqueue.on_readable", 0); -} - -void grpc_workqueue_enqueue(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, - grpc_closure *closure, grpc_error *error) { - GPR_TIMER_BEGIN("workqueue.enqueue", 0); - gpr_atm last = gpr_atm_full_fetch_add(&workqueue->state, 2); - GPR_ASSERT(last & 1); - closure->error = error; - gpr_mpscq_push(&workqueue->queue, &closure->next_data.atm_next); - if (last == 1) { - wakeup(exec_ctx, workqueue); - } - GPR_TIMER_END("workqueue.enqueue", 0); -} - -#endif /* GPR_POSIX_SOCKET */ diff --git a/src/core/lib/iomgr/workqueue_windows.c b/src/core/lib/iomgr/workqueue_windows.c index ee81dc248e..5c93d3c59e 100644 --- a/src/core/lib/iomgr/workqueue_windows.c +++ b/src/core/lib/iomgr/workqueue_windows.c @@ -43,12 +43,16 @@ // workqueues. #ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG -void grpc_workqueue_ref(grpc_workqueue *workqueue, const char *file, int line, - const char *reason) {} +grpc_workqueue *grpc_workqueue_ref(grpc_workqueue *workqueue, const char *file, + int line, const char *reason) { + return workqueue; +} void grpc_workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue, const char *file, int line, const char *reason) {} #else -void grpc_workqueue_ref(grpc_workqueue *workqueue) {} +grpc_workqueue *grpc_workqueue_ref(grpc_workqueue *workqueue) { + return workqueue; +} void grpc_workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue) {} #endif diff --git a/src/core/lib/profiling/basic_timers.c b/src/core/lib/profiling/basic_timers.c index 51813d0461..bdf9af2339 100644 --- a/src/core/lib/profiling/basic_timers.c +++ b/src/core/lib/profiling/basic_timers.c @@ -83,6 +83,7 @@ static int g_shutdown; static gpr_thd_id g_writing_thread; static __thread int g_thread_id; static int g_next_thread_id; +static int g_writing_enabled = 1; static int timer_log_push_back(gpr_timer_log_list *list, gpr_timer_log *log) { if (list->head == NULL) { @@ -177,7 +178,7 @@ static void flush_logs(gpr_timer_log_list *list) { } } -static void finish_writing() { +static void finish_writing(void) { pthread_mutex_lock(&g_mu); g_shutdown = 1; pthread_cond_signal(&g_cv); @@ -230,6 +231,10 @@ static void gpr_timers_log_add(const char *tagstr, marker_type type, int important, const char *file, int line) { gpr_timer_entry *entry; + if (!g_writing_enabled) { + return; + } + if (g_thread_log == NULL || g_thread_log->num_entries == MAX_COUNT) { rotate_log(); } @@ -261,6 +266,8 @@ void gpr_timer_end(const char *tagstr, int important, const char *file, gpr_timers_log_add(tagstr, END, important, file, line); } +void gpr_timer_set_enabled(int enabled) { g_writing_enabled = enabled; } + /* Basic profiler specific API functions. */ void gpr_timers_global_init(void) {} @@ -272,4 +279,6 @@ void gpr_timers_global_init(void) {} void gpr_timers_global_destroy(void) {} void gpr_timers_set_log_filename(const char *filename) {} + +void gpr_timer_set_enabled(int enabled) {} #endif /* GRPC_BASIC_PROFILER */ diff --git a/src/core/lib/profiling/timers.h b/src/core/lib/profiling/timers.h index c8567e8137..621cdbf656 100644 --- a/src/core/lib/profiling/timers.h +++ b/src/core/lib/profiling/timers.h @@ -50,6 +50,8 @@ void gpr_timer_end(const char *tagstr, int important, const char *file, void gpr_timers_set_log_filename(const char *filename); +void gpr_timer_set_enabled(int enabled); + #if !(defined(GRPC_STAP_PROFILER) + defined(GRPC_BASIC_PROFILER)) /* No profiling. No-op all the things. */ #define GPR_TIMER_MARK(tag, important) \ diff --git a/src/core/lib/support/log.c b/src/core/lib/support/log.c index 899f1218b6..af1651dae5 100644 --- a/src/core/lib/support/log.c +++ b/src/core/lib/support/log.c @@ -60,8 +60,9 @@ const char *gpr_log_severity_string(gpr_log_severity severity) { void gpr_log_message(const char *file, int line, gpr_log_severity severity, const char *message) { - if ((gpr_atm)severity < gpr_atm_no_barrier_load(&g_min_severity_to_print)) + if ((gpr_atm)severity < gpr_atm_no_barrier_load(&g_min_severity_to_print)) { return; + } gpr_log_func_args lfargs; memset(&lfargs, 0, sizeof(lfargs)); @@ -82,11 +83,11 @@ void gpr_log_verbosity_init() { gpr_atm min_severity_to_print = GPR_LOG_SEVERITY_ERROR; if (verbosity != NULL) { - if (strcmp(verbosity, "DEBUG") == 0) { + if (gpr_stricmp(verbosity, "DEBUG") == 0) { min_severity_to_print = (gpr_atm)GPR_LOG_SEVERITY_DEBUG; - } else if (strcmp(verbosity, "INFO") == 0) { + } else if (gpr_stricmp(verbosity, "INFO") == 0) { min_severity_to_print = (gpr_atm)GPR_LOG_SEVERITY_INFO; - } else if (strcmp(verbosity, "ERROR") == 0) { + } else if (gpr_stricmp(verbosity, "ERROR") == 0) { min_severity_to_print = (gpr_atm)GPR_LOG_SEVERITY_ERROR; } gpr_free(verbosity); diff --git a/src/core/lib/support/string.c b/src/core/lib/support/string.c index 30c1e67647..d17fb9da4b 100644 --- a/src/core/lib/support/string.c +++ b/src/core/lib/support/string.c @@ -304,3 +304,14 @@ void gpr_strvec_add(gpr_strvec *sv, char *str) { char *gpr_strvec_flatten(gpr_strvec *sv, size_t *final_length) { return gpr_strjoin((const char **)sv->strs, sv->count, final_length); } + +int gpr_stricmp(const char *a, const char *b) { + int ca, cb; + do { + ca = tolower(*a); + cb = tolower(*b); + ++a; + ++b; + } while (ca == cb && ca && cb); + return ca - cb; +} diff --git a/src/core/lib/support/string.h b/src/core/lib/support/string.h index 2b6bb3eec6..9a94e9471c 100644 --- a/src/core/lib/support/string.h +++ b/src/core/lib/support/string.h @@ -118,6 +118,10 @@ void gpr_strvec_add(gpr_strvec *strs, char *add); total_length as per gpr_strjoin */ char *gpr_strvec_flatten(gpr_strvec *strs, size_t *total_length); +/** Case insensitive string comparison... return <0 if lower(a)<lower(b), ==0 if + lower(a)==lower(b), >0 if lower(a)>lower(b) */ +int gpr_stricmp(const char *a, const char *b); + #ifdef __cplusplus } #endif diff --git a/src/core/lib/surface/call.c b/src/core/lib/surface/call.c index b0f66f4f61..30559304c6 100644 --- a/src/core/lib/surface/call.c +++ b/src/core/lib/surface/call.c @@ -223,33 +223,37 @@ static void destroy_call(grpc_exec_ctx *exec_ctx, void *call_stack, static void receiving_slice_ready(grpc_exec_ctx *exec_ctx, void *bctlp, grpc_error *error); -grpc_call *grpc_call_create( - grpc_channel *channel, grpc_call *parent_call, uint32_t propagation_mask, - grpc_completion_queue *cq, grpc_pollset_set *pollset_set_alternative, - const void *server_transport_data, grpc_mdelem **add_initial_metadata, - size_t add_initial_metadata_count, gpr_timespec send_deadline) { +grpc_error *grpc_call_create(const grpc_call_create_args *args, + grpc_call **out_call) { size_t i, j; - grpc_channel_stack *channel_stack = grpc_channel_get_channel_stack(channel); + grpc_channel_stack *channel_stack = + grpc_channel_get_channel_stack(args->channel); grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; grpc_call *call; GPR_TIMER_BEGIN("grpc_call_create", 0); call = gpr_malloc(sizeof(grpc_call) + channel_stack->call_stack_size); + *out_call = call; memset(call, 0, sizeof(grpc_call)); gpr_mu_init(&call->mu); - call->channel = channel; - call->cq = cq; - call->parent = parent_call; + call->channel = args->channel; + call->cq = args->cq; + call->parent = args->parent_call; /* Always support no compression */ GPR_BITSET(&call->encodings_accepted_by_peer, GRPC_COMPRESS_NONE); - call->is_client = server_transport_data == NULL; + call->is_client = args->server_transport_data == NULL; + grpc_mdstr *path = NULL; if (call->is_client) { - GPR_ASSERT(add_initial_metadata_count < MAX_SEND_EXTRA_METADATA_COUNT); - for (i = 0; i < add_initial_metadata_count; i++) { - call->send_extra_metadata[i].md = add_initial_metadata[i]; + GPR_ASSERT(args->add_initial_metadata_count < + MAX_SEND_EXTRA_METADATA_COUNT); + for (i = 0; i < args->add_initial_metadata_count; i++) { + call->send_extra_metadata[i].md = args->add_initial_metadata[i]; + if (args->add_initial_metadata[i]->key == GRPC_MDSTR_PATH) { + path = GRPC_MDSTR_REF(args->add_initial_metadata[i]->value); + } } - call->send_extra_metadata_count = (int)add_initial_metadata_count; + call->send_extra_metadata_count = (int)args->add_initial_metadata_count; } else { - GPR_ASSERT(add_initial_metadata_count == 0); + GPR_ASSERT(args->add_initial_metadata_count == 0); call->send_extra_metadata_count = 0; } for (i = 0; i < 2; i++) { @@ -257,84 +261,87 @@ grpc_call *grpc_call_create( call->metadata_batch[i][j].deadline = gpr_inf_future(GPR_CLOCK_MONOTONIC); } } - send_deadline = gpr_convert_clock_type(send_deadline, GPR_CLOCK_MONOTONIC); + gpr_timespec send_deadline = + gpr_convert_clock_type(args->send_deadline, GPR_CLOCK_MONOTONIC); - if (parent_call != NULL) { - GRPC_CALL_INTERNAL_REF(parent_call, "child"); + if (args->parent_call != NULL) { + GRPC_CALL_INTERNAL_REF(args->parent_call, "child"); GPR_ASSERT(call->is_client); - GPR_ASSERT(!parent_call->is_client); + GPR_ASSERT(!args->parent_call->is_client); - gpr_mu_lock(&parent_call->mu); + gpr_mu_lock(&args->parent_call->mu); - if (propagation_mask & GRPC_PROPAGATE_DEADLINE) { + if (args->propagation_mask & GRPC_PROPAGATE_DEADLINE) { send_deadline = gpr_time_min( gpr_convert_clock_type(send_deadline, - parent_call->send_deadline.clock_type), - parent_call->send_deadline); + args->parent_call->send_deadline.clock_type), + args->parent_call->send_deadline); } /* for now GRPC_PROPAGATE_TRACING_CONTEXT *MUST* be passed with * GRPC_PROPAGATE_STATS_CONTEXT */ /* TODO(ctiller): This should change to use the appropriate census start_op * call. */ - if (propagation_mask & GRPC_PROPAGATE_CENSUS_TRACING_CONTEXT) { - GPR_ASSERT(propagation_mask & GRPC_PROPAGATE_CENSUS_STATS_CONTEXT); - grpc_call_context_set(call, GRPC_CONTEXT_TRACING, - parent_call->context[GRPC_CONTEXT_TRACING].value, - NULL); + if (args->propagation_mask & GRPC_PROPAGATE_CENSUS_TRACING_CONTEXT) { + GPR_ASSERT(args->propagation_mask & GRPC_PROPAGATE_CENSUS_STATS_CONTEXT); + grpc_call_context_set( + call, GRPC_CONTEXT_TRACING, + args->parent_call->context[GRPC_CONTEXT_TRACING].value, NULL); } else { - GPR_ASSERT(propagation_mask & GRPC_PROPAGATE_CENSUS_STATS_CONTEXT); + GPR_ASSERT(args->propagation_mask & GRPC_PROPAGATE_CENSUS_STATS_CONTEXT); } - if (propagation_mask & GRPC_PROPAGATE_CANCELLATION) { + if (args->propagation_mask & GRPC_PROPAGATE_CANCELLATION) { call->cancellation_is_inherited = 1; } - if (parent_call->first_child == NULL) { - parent_call->first_child = call; + if (args->parent_call->first_child == NULL) { + args->parent_call->first_child = call; call->sibling_next = call->sibling_prev = call; } else { - call->sibling_next = parent_call->first_child; - call->sibling_prev = parent_call->first_child->sibling_prev; + call->sibling_next = args->parent_call->first_child; + call->sibling_prev = args->parent_call->first_child->sibling_prev; call->sibling_next->sibling_prev = call->sibling_prev->sibling_next = call; } - gpr_mu_unlock(&parent_call->mu); + gpr_mu_unlock(&args->parent_call->mu); } call->send_deadline = send_deadline; - GRPC_CHANNEL_INTERNAL_REF(channel, "call"); + GRPC_CHANNEL_INTERNAL_REF(args->channel, "call"); /* initial refcount dropped by grpc_call_destroy */ - grpc_error *error = grpc_call_stack_init( - &exec_ctx, channel_stack, 1, destroy_call, call, call->context, - server_transport_data, send_deadline, CALL_STACK_FROM_CALL(call)); + grpc_error *error = + grpc_call_stack_init(&exec_ctx, channel_stack, 1, destroy_call, call, + call->context, args->server_transport_data, path, + send_deadline, CALL_STACK_FROM_CALL(call)); if (error != GRPC_ERROR_NONE) { grpc_status_code status; const char *error_str; grpc_error_get_status(error, &status, &error_str); close_with_status(&exec_ctx, call, status, error_str); - GRPC_ERROR_UNREF(error); } - if (cq != NULL) { + if (args->cq != NULL) { GPR_ASSERT( - pollset_set_alternative == NULL && + args->pollset_set_alternative == NULL && "Only one of 'cq' and 'pollset_set_alternative' should be non-NULL."); - GRPC_CQ_INTERNAL_REF(cq, "bind"); + GRPC_CQ_INTERNAL_REF(args->cq, "bind"); call->pollent = - grpc_polling_entity_create_from_pollset(grpc_cq_pollset(cq)); + grpc_polling_entity_create_from_pollset(grpc_cq_pollset(args->cq)); } - if (pollset_set_alternative != NULL) { - call->pollent = - grpc_polling_entity_create_from_pollset_set(pollset_set_alternative); + if (args->pollset_set_alternative != NULL) { + call->pollent = grpc_polling_entity_create_from_pollset_set( + args->pollset_set_alternative); } if (!grpc_polling_entity_is_empty(&call->pollent)) { grpc_call_stack_set_pollset_or_pollset_set( &exec_ctx, CALL_STACK_FROM_CALL(call), &call->pollent); } + if (path != NULL) GRPC_MDSTR_UNREF(path); + grpc_exec_ctx_finish(&exec_ctx); GPR_TIMER_END("grpc_call_create", 0); - return call; + return error; } void grpc_call_set_completion_queue(grpc_exec_ctx *exec_ctx, grpc_call *call, @@ -1038,9 +1045,14 @@ static void finish_batch_completion(grpc_exec_ctx *exec_ctx, void *user_data, static void post_batch_completion(grpc_exec_ctx *exec_ctx, batch_control *bctl) { grpc_call *call = bctl->call; + grpc_error *error = bctl->error; + if (bctl->recv_final_op) { + GRPC_ERROR_UNREF(error); + error = GRPC_ERROR_NONE; + } if (bctl->is_notify_tag_closure) { /* unrefs bctl->error */ - grpc_exec_ctx_sched(exec_ctx, bctl->notify_tag, bctl->error, NULL); + grpc_closure_run(exec_ctx, bctl->notify_tag, error); gpr_mu_lock(&call->mu); bctl->call->used_batches = (uint8_t)(bctl->call->used_batches & @@ -1049,7 +1061,7 @@ static void post_batch_completion(grpc_exec_ctx *exec_ctx, GRPC_CALL_INTERNAL_UNREF(exec_ctx, call, "completion"); } else { /* unrefs bctl->error */ - grpc_cq_end_op(exec_ctx, bctl->call->cq, bctl->notify_tag, bctl->error, + grpc_cq_end_op(exec_ctx, bctl->call->cq, bctl->notify_tag, error, finish_batch_completion, bctl, &bctl->cq_completion); } } @@ -1198,6 +1210,14 @@ static void validate_filtered_metadata(grpc_exec_ctx *exec_ctx, } } +static void add_batch_error(batch_control *bctl, grpc_error *error) { + if (error == GRPC_ERROR_NONE) return; + if (bctl->error == GRPC_ERROR_NONE) { + bctl->error = GRPC_ERROR_CREATE("Call batch operation failed"); + } + bctl->error = grpc_error_add_child(bctl->error, error); +} + static void receiving_initial_metadata_ready(grpc_exec_ctx *exec_ctx, void *bctlp, grpc_error *error) { batch_control *bctl = bctlp; @@ -1205,9 +1225,8 @@ static void receiving_initial_metadata_ready(grpc_exec_ctx *exec_ctx, gpr_mu_lock(&call->mu); - if (error != GRPC_ERROR_NONE) { - bctl->error = GRPC_ERROR_REF(error); - } else { + add_batch_error(bctl, GRPC_ERROR_REF(error)); + if (error == GRPC_ERROR_NONE) { grpc_metadata_batch *md = &call->metadata_batch[1 /* is_receiving */][0 /* is_trailing */]; grpc_metadata_batch_filter(md, recv_initial_filter, call); @@ -1304,8 +1323,7 @@ static void finish_batch(grpc_exec_ctx *exec_ctx, void *bctlp, GRPC_ERROR_UNREF(error); error = GRPC_ERROR_NONE; } - GRPC_ERROR_UNREF(bctl->error); - bctl->error = GRPC_ERROR_REF(error); + add_batch_error(bctl, GRPC_ERROR_REF(error)); gpr_mu_unlock(&call->mu); if (gpr_unref(&bctl->steps_to_complete)) { post_batch_completion(exec_ctx, bctl); @@ -1341,6 +1359,7 @@ static grpc_call_error call_start_batch(grpc_exec_ctx *exec_ctx, grpc_transport_stream_op *stream_op = &bctl->op; memset(stream_op, 0, sizeof(*stream_op)); + stream_op->covered_by_poller = true; if (nops == 0) { GRPC_CALL_INTERNAL_REF(call, "completion"); diff --git a/src/core/lib/surface/call.h b/src/core/lib/surface/call.h index 3a78fe3aa3..18af41b7fb 100644 --- a/src/core/lib/surface/call.h +++ b/src/core/lib/surface/call.h @@ -49,15 +49,29 @@ typedef void (*grpc_ioreq_completion_func)(grpc_exec_ctx *exec_ctx, grpc_call *call, int success, void *user_data); -grpc_call *grpc_call_create(grpc_channel *channel, grpc_call *parent_call, - uint32_t propagation_mask, - grpc_completion_queue *cq, - /* if not NULL, it'll be used in lieu of \a cq */ - grpc_pollset_set *pollset_set_alternative, - const void *server_transport_data, - grpc_mdelem **add_initial_metadata, - size_t add_initial_metadata_count, - gpr_timespec send_deadline); +typedef struct grpc_call_create_args { + grpc_channel *channel; + + grpc_call *parent_call; + uint32_t propagation_mask; + + grpc_completion_queue *cq; + /* if not NULL, it'll be used in lieu of cq */ + grpc_pollset_set *pollset_set_alternative; + + const void *server_transport_data; + + grpc_mdelem **add_initial_metadata; + size_t add_initial_metadata_count; + + gpr_timespec send_deadline; +} grpc_call_create_args; + +/* Create a new call based on \a args. + Regardless of success or failure, always returns a valid new call into *call + */ +grpc_error *grpc_call_create(const grpc_call_create_args *args, + grpc_call **call); void grpc_call_set_completion_queue(grpc_exec_ctx *exec_ctx, grpc_call *call, grpc_completion_queue *cq); diff --git a/src/core/lib/surface/channel.c b/src/core/lib/surface/channel.c index 6adb70a987..92d783b78d 100644 --- a/src/core/lib/surface/channel.c +++ b/src/core/lib/surface/channel.c @@ -193,9 +193,21 @@ static grpc_call *grpc_channel_create_call_internal( send_metadata[num_metadata++] = GRPC_MDELEM_REF(channel->default_authority); } - return grpc_call_create(channel, parent_call, propagation_mask, cq, - pollset_set_alternative, NULL, send_metadata, - num_metadata, deadline); + grpc_call_create_args args; + memset(&args, 0, sizeof(args)); + args.channel = channel; + args.parent_call = parent_call; + args.propagation_mask = propagation_mask; + args.cq = cq; + args.pollset_set_alternative = pollset_set_alternative; + args.server_transport_data = NULL; + args.add_initial_metadata = send_metadata; + args.add_initial_metadata_count = num_metadata; + args.send_deadline = deadline; + + grpc_call *call; + GRPC_LOG_IF_ERROR("call_create", grpc_call_create(&args, &call)); + return call; } grpc_call *grpc_channel_create_call(grpc_channel *channel, diff --git a/src/core/lib/surface/completion_queue.c b/src/core/lib/surface/completion_queue.c index 5978884db8..4e0feb56ac 100644 --- a/src/core/lib/surface/completion_queue.c +++ b/src/core/lib/surface/completion_queue.c @@ -39,6 +39,7 @@ #include <grpc/support/alloc.h> #include <grpc/support/atm.h> #include <grpc/support/log.h> +#include <grpc/support/string_util.h> #include <grpc/support/time.h> #include "src/core/lib/iomgr/pollset.h" @@ -50,6 +51,9 @@ #include "src/core/lib/surface/event_string.h" int grpc_trace_operation_failures; +#ifndef NDEBUG +int grpc_trace_pending_tags; +#endif typedef struct { grpc_pollset_worker **worker; @@ -67,6 +71,9 @@ struct grpc_completion_queue { gpr_refcount pending_events; /** Once owning_refs drops to zero, we will destroy the cq */ gpr_refcount owning_refs; + /** counter of how many things have ever been queued on this completion queue + useful for avoiding locks to check the queue */ + gpr_atm things_queued_ever; /** 0 initially, 1 once we've begun shutting down */ int shutdown; int shutdown_called; @@ -121,15 +128,6 @@ void grpc_cq_global_shutdown(void) { } } -struct grpc_cq_alarm { - grpc_timer alarm; - grpc_cq_completion completion; - /** completion queue where events about this alarm will be posted */ - grpc_completion_queue *cq; - /** user supplied tag */ - void *tag; -}; - grpc_completion_queue *grpc_completion_queue_create(void *reserved) { grpc_completion_queue *cc; GPR_ASSERT(!reserved); @@ -166,6 +164,7 @@ grpc_completion_queue *grpc_completion_queue_create(void *reserved) { cc->is_server_cq = 0; cc->is_non_listening_server_cq = 0; cc->num_pluckers = 0; + gpr_atm_no_barrier_store(&cc->things_queued_ever, 0); #ifndef NDEBUG cc->outstanding_tag_count = 0; #endif @@ -276,6 +275,7 @@ void grpc_cq_end_op(grpc_exec_ctx *exec_ctx, grpc_completion_queue *cc, GPR_ASSERT(found); #endif shutdown = gpr_unref(&cc->pending_events); + gpr_atm_no_barrier_fetch_add(&cc->things_queued_ever, 1); if (!shutdown) { cc->completed_tail->next = ((uintptr_t)storage) | (1u & (uintptr_t)cc->completed_tail->next); @@ -313,13 +313,66 @@ void grpc_cq_end_op(grpc_exec_ctx *exec_ctx, grpc_completion_queue *cc, GRPC_ERROR_UNREF(error); } +typedef struct { + gpr_atm last_seen_things_queued_ever; + grpc_completion_queue *cq; + gpr_timespec deadline; + grpc_cq_completion *stolen_completion; + void *tag; /* for pluck */ + bool first_loop; +} cq_is_finished_arg; + +static bool cq_is_next_finished(grpc_exec_ctx *exec_ctx, void *arg) { + cq_is_finished_arg *a = arg; + grpc_completion_queue *cq = a->cq; + GPR_ASSERT(a->stolen_completion == NULL); + gpr_atm current_last_seen_things_queued_ever = + gpr_atm_no_barrier_load(&cq->things_queued_ever); + if (current_last_seen_things_queued_ever != a->last_seen_things_queued_ever) { + gpr_mu_lock(cq->mu); + a->last_seen_things_queued_ever = + gpr_atm_no_barrier_load(&cq->things_queued_ever); + if (cq->completed_tail != &cq->completed_head) { + a->stolen_completion = (grpc_cq_completion *)cq->completed_head.next; + cq->completed_head.next = a->stolen_completion->next & ~(uintptr_t)1; + if (a->stolen_completion == cq->completed_tail) { + cq->completed_tail = &cq->completed_head; + } + gpr_mu_unlock(cq->mu); + return true; + } + gpr_mu_unlock(cq->mu); + } + return !a->first_loop && + gpr_time_cmp(a->deadline, gpr_now(a->deadline.clock_type)) < 0; +} + +#ifndef NDEBUG +static void dump_pending_tags(grpc_completion_queue *cc) { + if (!grpc_trace_pending_tags) return; + + gpr_strvec v; + gpr_strvec_init(&v); + gpr_strvec_add(&v, gpr_strdup("PENDING TAGS:")); + for (size_t i = 0; i < cc->outstanding_tag_count; i++) { + char *s; + gpr_asprintf(&s, " %p", cc->outstanding_tags[i]); + gpr_strvec_add(&v, s); + } + char *out = gpr_strvec_flatten(&v, NULL); + gpr_strvec_destroy(&v); + gpr_log(GPR_DEBUG, "%s", out); + gpr_free(out); +} +#else +static void dump_pending_tags(grpc_completion_queue *cc) {} +#endif + grpc_event grpc_completion_queue_next(grpc_completion_queue *cc, gpr_timespec deadline, void *reserved) { grpc_event ret; grpc_pollset_worker *worker = NULL; - int first_loop = 1; gpr_timespec now; - grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; GPR_TIMER_BEGIN("grpc_completion_queue_next", 0); @@ -333,11 +386,33 @@ grpc_event grpc_completion_queue_next(grpc_completion_queue *cc, reserved)); GPR_ASSERT(!reserved); + dump_pending_tags(cc); + deadline = gpr_convert_clock_type(deadline, GPR_CLOCK_MONOTONIC); GRPC_CQ_INTERNAL_REF(cc, "next"); gpr_mu_lock(cc->mu); + cq_is_finished_arg is_finished_arg = { + .last_seen_things_queued_ever = + gpr_atm_no_barrier_load(&cc->things_queued_ever), + .cq = cc, + .deadline = deadline, + .stolen_completion = NULL, + .tag = NULL, + .first_loop = true}; + grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT_WITH_FINISH_CHECK( + cq_is_next_finished, &is_finished_arg); for (;;) { + if (is_finished_arg.stolen_completion != NULL) { + gpr_mu_unlock(cc->mu); + grpc_cq_completion *c = is_finished_arg.stolen_completion; + is_finished_arg.stolen_completion = NULL; + ret.type = GRPC_OP_COMPLETE; + ret.success = c->next & 1u; + ret.tag = c->tag; + c->done(&exec_ctx, c->done_arg, c); + break; + } if (cc->completed_tail != &cc->completed_head) { grpc_cq_completion *c = (grpc_cq_completion *)cc->completed_head.next; cc->completed_head.next = c->next & ~(uintptr_t)1; @@ -358,13 +433,13 @@ grpc_event grpc_completion_queue_next(grpc_completion_queue *cc, break; } now = gpr_now(GPR_CLOCK_MONOTONIC); - if (!first_loop && gpr_time_cmp(now, deadline) >= 0) { + if (!is_finished_arg.first_loop && gpr_time_cmp(now, deadline) >= 0) { gpr_mu_unlock(cc->mu); memset(&ret, 0, sizeof(ret)); ret.type = GRPC_QUEUE_TIMEOUT; + dump_pending_tags(cc); break; } - first_loop = 0; /* Check alarms - these are a global resource so we just ping each time through on every pollset. May update deadline to ensure timely wakeups. @@ -387,13 +462,16 @@ grpc_event grpc_completion_queue_next(grpc_completion_queue *cc, GRPC_ERROR_UNREF(err); memset(&ret, 0, sizeof(ret)); ret.type = GRPC_QUEUE_TIMEOUT; + dump_pending_tags(cc); break; } } + is_finished_arg.first_loop = false; } GRPC_SURFACE_TRACE_RETURNED_EVENT(cc, &ret); GRPC_CQ_INTERNAL_UNREF(cc, "next"); grpc_exec_ctx_finish(&exec_ctx); + GPR_ASSERT(is_finished_arg.stolen_completion == NULL); GPR_TIMER_END("grpc_completion_queue_next", 0); @@ -424,6 +502,37 @@ static void del_plucker(grpc_completion_queue *cc, void *tag, GPR_UNREACHABLE_CODE(return ); } +static bool cq_is_pluck_finished(grpc_exec_ctx *exec_ctx, void *arg) { + cq_is_finished_arg *a = arg; + grpc_completion_queue *cq = a->cq; + GPR_ASSERT(a->stolen_completion == NULL); + gpr_atm current_last_seen_things_queued_ever = + gpr_atm_no_barrier_load(&cq->things_queued_ever); + if (current_last_seen_things_queued_ever != a->last_seen_things_queued_ever) { + gpr_mu_lock(cq->mu); + a->last_seen_things_queued_ever = + gpr_atm_no_barrier_load(&cq->things_queued_ever); + grpc_cq_completion *c; + grpc_cq_completion *prev = &cq->completed_head; + while ((c = (grpc_cq_completion *)(prev->next & ~(uintptr_t)1)) != + &cq->completed_head) { + if (c->tag == a->tag) { + prev->next = (prev->next & (uintptr_t)1) | (c->next & ~(uintptr_t)1); + if (c == cq->completed_tail) { + cq->completed_tail = prev; + } + gpr_mu_unlock(cq->mu); + a->stolen_completion = c; + return true; + } + prev = c; + } + gpr_mu_unlock(cq->mu); + } + return !a->first_loop && + gpr_time_cmp(a->deadline, gpr_now(a->deadline.clock_type)) < 0; +} + grpc_event grpc_completion_queue_pluck(grpc_completion_queue *cc, void *tag, gpr_timespec deadline, void *reserved) { grpc_event ret; @@ -431,8 +540,6 @@ grpc_event grpc_completion_queue_pluck(grpc_completion_queue *cc, void *tag, grpc_cq_completion *prev; grpc_pollset_worker *worker = NULL; gpr_timespec now; - int first_loop = 1; - grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT; GPR_TIMER_BEGIN("grpc_completion_queue_pluck", 0); @@ -448,11 +555,33 @@ grpc_event grpc_completion_queue_pluck(grpc_completion_queue *cc, void *tag, } GPR_ASSERT(!reserved); + dump_pending_tags(cc); + deadline = gpr_convert_clock_type(deadline, GPR_CLOCK_MONOTONIC); GRPC_CQ_INTERNAL_REF(cc, "pluck"); gpr_mu_lock(cc->mu); + cq_is_finished_arg is_finished_arg = { + .last_seen_things_queued_ever = + gpr_atm_no_barrier_load(&cc->things_queued_ever), + .cq = cc, + .deadline = deadline, + .stolen_completion = NULL, + .tag = tag, + .first_loop = true}; + grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT_WITH_FINISH_CHECK( + cq_is_pluck_finished, &is_finished_arg); for (;;) { + if (is_finished_arg.stolen_completion != NULL) { + gpr_mu_unlock(cc->mu); + c = is_finished_arg.stolen_completion; + is_finished_arg.stolen_completion = NULL; + ret.type = GRPC_OP_COMPLETE; + ret.success = c->next & 1u; + ret.tag = c->tag; + c->done(&exec_ctx, c->done_arg, c); + break; + } prev = &cc->completed_head; while ((c = (grpc_cq_completion *)(prev->next & ~(uintptr_t)1)) != &cc->completed_head) { @@ -485,17 +614,18 @@ grpc_event grpc_completion_queue_pluck(grpc_completion_queue *cc, void *tag, memset(&ret, 0, sizeof(ret)); /* TODO(ctiller): should we use a different result here */ ret.type = GRPC_QUEUE_TIMEOUT; + dump_pending_tags(cc); break; } now = gpr_now(GPR_CLOCK_MONOTONIC); - if (!first_loop && gpr_time_cmp(now, deadline) >= 0) { + if (!is_finished_arg.first_loop && gpr_time_cmp(now, deadline) >= 0) { del_plucker(cc, tag, &worker); gpr_mu_unlock(cc->mu); memset(&ret, 0, sizeof(ret)); ret.type = GRPC_QUEUE_TIMEOUT; + dump_pending_tags(cc); break; } - first_loop = 0; /* Check alarms - these are a global resource so we just ping each time through on every pollset. May update deadline to ensure timely wakeups. @@ -518,15 +648,18 @@ grpc_event grpc_completion_queue_pluck(grpc_completion_queue *cc, void *tag, GRPC_ERROR_UNREF(err); memset(&ret, 0, sizeof(ret)); ret.type = GRPC_QUEUE_TIMEOUT; + dump_pending_tags(cc); break; } } + is_finished_arg.first_loop = false; del_plucker(cc, tag, &worker); } done: GRPC_SURFACE_TRACE_RETURNED_EVENT(cc, &ret); GRPC_CQ_INTERNAL_UNREF(cc, "pluck"); grpc_exec_ctx_finish(&exec_ctx); + GPR_ASSERT(is_finished_arg.stolen_completion == NULL); GPR_TIMER_END("grpc_completion_queue_pluck", 0); diff --git a/src/core/lib/surface/completion_queue.h b/src/core/lib/surface/completion_queue.h index 4dbf3aae63..c1cafba5f2 100644 --- a/src/core/lib/surface/completion_queue.h +++ b/src/core/lib/surface/completion_queue.h @@ -44,6 +44,9 @@ extern int grpc_cq_pluck_trace; extern int grpc_cq_event_timeout_trace; extern int grpc_trace_operation_failures; +#ifndef NDEBUG +extern int grpc_trace_pending_tags; +#endif typedef struct grpc_cq_completion { /** user supplied tag */ diff --git a/src/core/lib/surface/init.c b/src/core/lib/surface/init.c index 289f4ce8e8..8ca0643ba7 100644 --- a/src/core/lib/surface/init.c +++ b/src/core/lib/surface/init.c @@ -184,12 +184,16 @@ void grpc_init(void) { grpc_register_tracer("compression", &grpc_compression_trace); grpc_register_tracer("queue_pluck", &grpc_cq_pluck_trace); grpc_register_tracer("combiner", &grpc_combiner_trace); + grpc_register_tracer("server_channel", &grpc_server_channel_trace); // Default pluck trace to 1 grpc_cq_pluck_trace = 1; grpc_register_tracer("queue_timeout", &grpc_cq_event_timeout_trace); // Default timeout trace to 1 grpc_cq_event_timeout_trace = 1; grpc_register_tracer("op_failure", &grpc_trace_operation_failures); +#ifndef NDEBUG + grpc_register_tracer("pending_tags", &grpc_trace_pending_tags); +#endif grpc_security_pre_init(); grpc_iomgr_init(); grpc_executor_init(); diff --git a/src/core/lib/surface/server.c b/src/core/lib/surface/server.c index 7300d79b9f..3a90308058 100644 --- a/src/core/lib/surface/server.c +++ b/src/core/lib/surface/server.c @@ -71,6 +71,8 @@ typedef struct registered_method registered_method; typedef enum { BATCH_CALL, REGISTERED_CALL } requested_call_type; +int grpc_server_channel_trace = 0; + typedef struct requested_call { requested_call_type type; size_t cq_idx; @@ -280,6 +282,7 @@ static void send_shutdown(grpc_exec_ctx *exec_ctx, grpc_channel *channel, grpc_channel_element *elem; op->send_goaway = send_goaway; + op->set_accept_stream = true; sc->slice = gpr_slice_from_copied_string("Server shutdown"); op->goaway_message = &sc->slice; op->goaway_status = GRPC_STATUS_OK; @@ -439,6 +442,13 @@ static void destroy_channel(grpc_exec_ctx *exec_ctx, channel_data *chand, chand->finish_destroy_channel_closure.cb = finish_destroy_channel; chand->finish_destroy_channel_closure.cb_arg = chand; + if (grpc_server_channel_trace && error != GRPC_ERROR_NONE) { + const char *msg = grpc_error_string(error); + gpr_log(GPR_INFO, "Disconnected client: %s", msg); + grpc_error_free_string(msg); + } + GRPC_ERROR_UNREF(error); + grpc_transport_op *op = grpc_make_transport_op(&chand->finish_destroy_channel_closure); op->set_accept_stream = true; @@ -446,13 +456,6 @@ static void destroy_channel(grpc_exec_ctx *exec_ctx, channel_data *chand, grpc_channel_stack_element( grpc_channel_get_channel_stack(chand->channel), 0), op); - - if (error != GRPC_ERROR_NONE) { - const char *msg = grpc_error_string(error); - gpr_log(GPR_INFO, "Disconnected client: %s", msg); - grpc_error_free_string(msg); - } - GRPC_ERROR_UNREF(error); } static void cpstr(char **dest, size_t *capacity, grpc_mdstr *value) { @@ -773,8 +776,7 @@ static void server_on_recv_initial_metadata(grpc_exec_ctx *exec_ctx, void *ptr, GRPC_ERROR_CREATE_REFERENCING("Missing :authority or :path", &error, 1); } - grpc_exec_ctx_sched(exec_ctx, calld->on_done_recv_initial_metadata, error, - NULL); + grpc_closure_run(exec_ctx, calld->on_done_recv_initial_metadata, error); } static void server_mutate_op(grpc_call_element *elem, @@ -829,11 +831,20 @@ static void accept_stream(grpc_exec_ctx *exec_ctx, void *cd, const void *transport_server_data) { channel_data *chand = cd; /* create a call */ - grpc_call *call = grpc_call_create(chand->channel, NULL, 0, NULL, NULL, - transport_server_data, NULL, 0, - gpr_inf_future(GPR_CLOCK_MONOTONIC)); + grpc_call_create_args args; + memset(&args, 0, sizeof(args)); + args.channel = chand->channel; + args.server_transport_data = transport_server_data; + args.send_deadline = gpr_inf_future(GPR_CLOCK_MONOTONIC); + grpc_call *call; + grpc_error *error = grpc_call_create(&args, &call); grpc_call_element *elem = grpc_call_stack_element(grpc_call_get_call_stack(call), 0); + if (error != GRPC_ERROR_NONE) { + got_initial_metadata(exec_ctx, elem, error); + GRPC_ERROR_UNREF(error); + return; + } call_data *calld = elem->call_data; grpc_op op; memset(&op, 0, sizeof(op)); diff --git a/src/core/lib/surface/server.h b/src/core/lib/surface/server.h index 551a40a4ff..a85d9f4964 100644 --- a/src/core/lib/surface/server.h +++ b/src/core/lib/surface/server.h @@ -40,6 +40,9 @@ extern const grpc_channel_filter grpc_server_top_filter; +/** Lightweight tracing of server channel state */ +extern int grpc_server_channel_trace; + /* Add a listener to the server: when the server starts, it will call start, and when it shuts down, it will call destroy */ void grpc_server_add_listener( diff --git a/src/core/lib/transport/connectivity_state.c b/src/core/lib/transport/connectivity_state.c index 68d05e3a85..fdb5307814 100644 --- a/src/core/lib/transport/connectivity_state.c +++ b/src/core/lib/transport/connectivity_state.c @@ -180,7 +180,8 @@ void grpc_connectivity_state_set(grpc_exec_ctx *exec_ctx, *w->current = tracker->current_state; tracker->watchers = w->next; if (grpc_connectivity_state_trace) { - gpr_log(GPR_DEBUG, "NOTIFY: %p", w->notify); + gpr_log(GPR_DEBUG, "NOTIFY: %p %s: %p", tracker, tracker->name, + w->notify); } grpc_exec_ctx_sched(exec_ctx, w->notify, GRPC_ERROR_REF(tracker->current_error), NULL); diff --git a/src/core/lib/transport/mdstr_hash_table.c b/src/core/lib/transport/mdstr_hash_table.c new file mode 100644 index 0000000000..8e914c420b --- /dev/null +++ b/src/core/lib/transport/mdstr_hash_table.c @@ -0,0 +1,157 @@ +// +// Copyright 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 "src/core/lib/transport/mdstr_hash_table.h" + +#include <stdbool.h> +#include <string.h> + +#include <grpc/support/alloc.h> +#include <grpc/support/log.h> + +#include "src/core/lib/transport/metadata.h" + +struct grpc_mdstr_hash_table { + gpr_refcount refs; + size_t num_entries; + size_t size; + grpc_mdstr_hash_table_entry* entries; +}; + +// Helper function for insert and get operations that performs quadratic +// probing (https://en.wikipedia.org/wiki/Quadratic_probing). +static size_t grpc_mdstr_hash_table_find_index( + const grpc_mdstr_hash_table* table, const grpc_mdstr* key, + bool find_empty) { + for (size_t i = 0; i < table->size; ++i) { + const size_t idx = (key->hash + i * i) % table->size; + if (table->entries[idx].key == NULL) return find_empty ? idx : table->size; + if (table->entries[idx].key == key) return idx; + } + return table->size; // Not found. +} + +static void grpc_mdstr_hash_table_add( + grpc_mdstr_hash_table* table, grpc_mdstr* key, void* value, + const grpc_mdstr_hash_table_vtable* vtable) { + GPR_ASSERT(value != NULL); + const size_t idx = + grpc_mdstr_hash_table_find_index(table, key, true /* find_empty */); + GPR_ASSERT(idx != table->size); // Table should never be full. + grpc_mdstr_hash_table_entry* entry = &table->entries[idx]; + entry->key = GRPC_MDSTR_REF(key); + entry->value = vtable->copy_value(value); + entry->vtable = vtable; +} + +grpc_mdstr_hash_table* grpc_mdstr_hash_table_create( + size_t num_entries, grpc_mdstr_hash_table_entry* entries) { + grpc_mdstr_hash_table* table = gpr_malloc(sizeof(*table)); + memset(table, 0, sizeof(*table)); + gpr_ref_init(&table->refs, 1); + table->num_entries = num_entries; + // Quadratic probing gets best performance when the table is no more + // than half full. + table->size = num_entries * 2; + const size_t entry_size = sizeof(grpc_mdstr_hash_table_entry) * table->size; + table->entries = gpr_malloc(entry_size); + memset(table->entries, 0, entry_size); + for (size_t i = 0; i < num_entries; ++i) { + grpc_mdstr_hash_table_entry* entry = &entries[i]; + grpc_mdstr_hash_table_add(table, entry->key, entry->value, entry->vtable); + } + return table; +} + +grpc_mdstr_hash_table* grpc_mdstr_hash_table_ref(grpc_mdstr_hash_table* table) { + if (table != NULL) gpr_ref(&table->refs); + return table; +} + +int grpc_mdstr_hash_table_unref(grpc_mdstr_hash_table* table) { + if (table != NULL && gpr_unref(&table->refs)) { + for (size_t i = 0; i < table->size; ++i) { + grpc_mdstr_hash_table_entry* entry = &table->entries[i]; + if (entry->key != NULL) { + GRPC_MDSTR_UNREF(entry->key); + entry->vtable->destroy_value(entry->value); + } + } + gpr_free(table->entries); + gpr_free(table); + return 1; + } + return 0; +} + +size_t grpc_mdstr_hash_table_num_entries(const grpc_mdstr_hash_table* table) { + return table->num_entries; +} + +void* grpc_mdstr_hash_table_get(const grpc_mdstr_hash_table* table, + const grpc_mdstr* key) { + const size_t idx = + grpc_mdstr_hash_table_find_index(table, key, false /* find_empty */); + if (idx == table->size) return NULL; // Not found. + return table->entries[idx].value; +} + +int grpc_mdstr_hash_table_cmp(const grpc_mdstr_hash_table* table1, + const grpc_mdstr_hash_table* table2) { + // Compare by num_entries. + if (table1->num_entries < table2->num_entries) return -1; + if (table1->num_entries > table2->num_entries) return 1; + for (size_t i = 0; i < table1->num_entries; ++i) { + grpc_mdstr_hash_table_entry* e1 = &table1->entries[i]; + grpc_mdstr_hash_table_entry* e2 = &table2->entries[i]; + // Compare keys by hash value. + if (e1->key->hash < e2->key->hash) return -1; + if (e1->key->hash > e2->key->hash) return 1; + // Compare by vtable (pointer equality). + if (e1->vtable < e2->vtable) return -1; + if (e1->vtable > e2->vtable) return 1; + // Compare values via vtable. + const int value_result = e1->vtable->compare_value(e1->value, e2->value); + if (value_result != 0) return value_result; + } + return 0; +} + +void grpc_mdstr_hash_table_iterate( + const grpc_mdstr_hash_table* table, + void (*func)(const grpc_mdstr_hash_table_entry* entry, void* user_data), + void* user_data) { + for (size_t i = 0; i < table->size; ++i) { + if (table->entries[i].key != NULL) { + func(&table->entries[i], user_data); + } + } +} diff --git a/src/core/lib/transport/mdstr_hash_table.h b/src/core/lib/transport/mdstr_hash_table.h new file mode 100644 index 0000000000..bceb4df93d --- /dev/null +++ b/src/core/lib/transport/mdstr_hash_table.h @@ -0,0 +1,92 @@ +/* + * Copyright 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_TRANSPORT_MDSTR_HASH_TABLE_H +#define GRPC_CORE_LIB_TRANSPORT_MDSTR_HASH_TABLE_H + +#include "src/core/lib/transport/metadata.h" + +/** Hash table implementation. + * + * This implementation uses open addressing + * (https://en.wikipedia.org/wiki/Open_addressing) with quadratic + * probing (https://en.wikipedia.org/wiki/Quadratic_probing). + * + * The keys are \a grpc_mdstr objects. The values are arbitrary pointers + * with a common vtable. + * + * Hash tables are intentionally immutable, to avoid the need for locking. + */ + +typedef struct grpc_mdstr_hash_table grpc_mdstr_hash_table; + +typedef struct grpc_mdstr_hash_table_vtable { + void (*destroy_value)(void* value); + void* (*copy_value)(void* value); + int (*compare_value)(void* value1, void* value2); +} grpc_mdstr_hash_table_vtable; + +typedef struct grpc_mdstr_hash_table_entry { + grpc_mdstr* key; + void* value; /* Must not be NULL. */ + const grpc_mdstr_hash_table_vtable* vtable; +} grpc_mdstr_hash_table_entry; + +/** Creates a new hash table of containing \a entries, which is an array + of length \a num_entries. + Creates its own copy of all keys and values from \a entries. */ +grpc_mdstr_hash_table* grpc_mdstr_hash_table_create( + size_t num_entries, grpc_mdstr_hash_table_entry* entries); + +grpc_mdstr_hash_table* grpc_mdstr_hash_table_ref(grpc_mdstr_hash_table* table); +/** Returns 1 when \a table is destroyed. */ +int grpc_mdstr_hash_table_unref(grpc_mdstr_hash_table* table); + +/** Returns the number of entries in \a table. */ +size_t grpc_mdstr_hash_table_num_entries(const grpc_mdstr_hash_table* table); + +/** Returns the value from \a table associated with \a key. + Returns NULL if \a key is not found. */ +void* grpc_mdstr_hash_table_get(const grpc_mdstr_hash_table* table, + const grpc_mdstr* key); + +/** Compares two hash tables. + The sort order is stable but undefined. */ +int grpc_mdstr_hash_table_cmp(const grpc_mdstr_hash_table* table1, + const grpc_mdstr_hash_table* table2); + +/** Iterates over the entries in \a table, calling \a func for each entry. */ +void grpc_mdstr_hash_table_iterate( + const grpc_mdstr_hash_table* table, + void (*func)(const grpc_mdstr_hash_table_entry* entry, void* user_data), + void* user_data); + +#endif /* GRPC_CORE_LIB_TRANSPORT_MDSTR_HASH_TABLE_H */ diff --git a/src/core/lib/transport/static_metadata.c b/src/core/lib/transport/static_metadata.c index 5e0352a467..8b22592b45 100644 --- a/src/core/lib/transport/static_metadata.c +++ b/src/core/lib/transport/static_metadata.c @@ -126,9 +126,9 @@ const char *const grpc_static_metadata_strings[GRPC_STATIC_MDSTR_COUNT] = { "if-range", "if-unmodified-since", "last-modified", + "lb-cost-bin", + "lb-token", "link", - "load-reporting-initial", - "load-reporting-trailing", "location", "max-forwards", ":method", diff --git a/src/core/lib/transport/static_metadata.h b/src/core/lib/transport/static_metadata.h index 5b9ee1a60a..28ad6f2961 100644 --- a/src/core/lib/transport/static_metadata.h +++ b/src/core/lib/transport/static_metadata.h @@ -175,12 +175,12 @@ extern grpc_mdstr grpc_static_mdstr_table[GRPC_STATIC_MDSTR_COUNT]; #define GRPC_MDSTR_IF_UNMODIFIED_SINCE (&grpc_static_mdstr_table[62]) /* "last-modified" */ #define GRPC_MDSTR_LAST_MODIFIED (&grpc_static_mdstr_table[63]) +/* "lb-cost-bin" */ +#define GRPC_MDSTR_LB_COST_BIN (&grpc_static_mdstr_table[64]) +/* "lb-token" */ +#define GRPC_MDSTR_LB_TOKEN (&grpc_static_mdstr_table[65]) /* "link" */ -#define GRPC_MDSTR_LINK (&grpc_static_mdstr_table[64]) -/* "load-reporting-initial" */ -#define GRPC_MDSTR_LOAD_REPORTING_INITIAL (&grpc_static_mdstr_table[65]) -/* "load-reporting-trailing" */ -#define GRPC_MDSTR_LOAD_REPORTING_TRAILING (&grpc_static_mdstr_table[66]) +#define GRPC_MDSTR_LINK (&grpc_static_mdstr_table[66]) /* "location" */ #define GRPC_MDSTR_LOCATION (&grpc_static_mdstr_table[67]) /* "max-forwards" */ @@ -337,13 +337,12 @@ extern uintptr_t grpc_static_mdelem_user_data[GRPC_STATIC_MDELEM_COUNT]; #define GRPC_MDELEM_IF_UNMODIFIED_SINCE_EMPTY (&grpc_static_mdelem_table[44]) /* "last-modified": "" */ #define GRPC_MDELEM_LAST_MODIFIED_EMPTY (&grpc_static_mdelem_table[45]) +/* "lb-cost-bin": "" */ +#define GRPC_MDELEM_LB_COST_BIN_EMPTY (&grpc_static_mdelem_table[46]) +/* "lb-token": "" */ +#define GRPC_MDELEM_LB_TOKEN_EMPTY (&grpc_static_mdelem_table[47]) /* "link": "" */ -#define GRPC_MDELEM_LINK_EMPTY (&grpc_static_mdelem_table[46]) -/* "load-reporting-initial": "" */ -#define GRPC_MDELEM_LOAD_REPORTING_INITIAL_EMPTY (&grpc_static_mdelem_table[47]) -/* "load-reporting-trailing": "" */ -#define GRPC_MDELEM_LOAD_REPORTING_TRAILING_EMPTY \ - (&grpc_static_mdelem_table[48]) +#define GRPC_MDELEM_LINK_EMPTY (&grpc_static_mdelem_table[48]) /* "location": "" */ #define GRPC_MDELEM_LOCATION_EMPTY (&grpc_static_mdelem_table[49]) /* "max-forwards": "" */ diff --git a/src/core/lib/transport/transport.c b/src/core/lib/transport/transport.c index 82fc605218..75aec7a5b4 100644 --- a/src/core/lib/transport/transport.c +++ b/src/core/lib/transport/transport.c @@ -46,8 +46,9 @@ #ifdef GRPC_STREAM_REFCOUNT_DEBUG void grpc_stream_ref(grpc_stream_refcount *refcount, const char *reason) { gpr_atm val = gpr_atm_no_barrier_load(&refcount->refs.count); - gpr_log(GPR_DEBUG, "%s %p:%p REF %d->%d %s", refcount->object_type, - refcount, refcount->destroy.cb_arg, (int)val, (int)val + 1, reason); + gpr_log(GPR_DEBUG, "%s %p:%p REF %" PRIdPTR "->%" PRIdPTR " %s", + refcount->object_type, refcount, refcount->destroy.cb_arg, val, + val + 1, reason); #else void grpc_stream_ref(grpc_stream_refcount *refcount) { #endif @@ -58,8 +59,9 @@ void grpc_stream_ref(grpc_stream_refcount *refcount) { void grpc_stream_unref(grpc_exec_ctx *exec_ctx, grpc_stream_refcount *refcount, const char *reason) { gpr_atm val = gpr_atm_no_barrier_load(&refcount->refs.count); - gpr_log(GPR_DEBUG, "%s %p:%p UNREF %d->%d %s", refcount->object_type, - refcount, refcount->destroy.cb_arg, (int)val, (int)val - 1, reason); + gpr_log(GPR_DEBUG, "%s %p:%p UNREF %" PRIdPTR "->%" PRIdPTR " %s", + refcount->object_type, refcount, refcount->destroy.cb_arg, val, + val - 1, reason); #else void grpc_stream_unref(grpc_exec_ctx *exec_ctx, grpc_stream_refcount *refcount) { @@ -274,3 +276,28 @@ grpc_transport_op *grpc_make_transport_op(grpc_closure *on_complete) { op->op.on_consumed = &op->outer_on_complete; return &op->op; } + +typedef struct { + grpc_closure outer_on_complete; + grpc_closure *inner_on_complete; + grpc_transport_stream_op op; +} made_transport_stream_op; + +static void destroy_made_transport_stream_op(grpc_exec_ctx *exec_ctx, void *arg, + grpc_error *error) { + made_transport_stream_op *op = arg; + grpc_exec_ctx_sched(exec_ctx, op->inner_on_complete, GRPC_ERROR_REF(error), + NULL); + gpr_free(op); +} + +grpc_transport_stream_op *grpc_make_transport_stream_op( + grpc_closure *on_complete) { + made_transport_stream_op *op = gpr_malloc(sizeof(*op)); + grpc_closure_init(&op->outer_on_complete, destroy_made_transport_stream_op, + op); + op->inner_on_complete = on_complete; + memset(&op->op, 0, sizeof(op->op)); + op->op.on_complete = &op->outer_on_complete; + return &op->op; +} diff --git a/src/core/lib/transport/transport.h b/src/core/lib/transport/transport.h index 8dc393fd61..50253ebad1 100644 --- a/src/core/lib/transport/transport.h +++ b/src/core/lib/transport/transport.h @@ -113,6 +113,10 @@ typedef struct grpc_transport_stream_op { have been completed. */ grpc_closure *on_complete; + /** Is the completion of this op covered by a poller (if false: the op should + complete independently of some pollset being polled) */ + bool covered_by_poller; + /** Send initial metadata to the peer, from the provided metadata batch. idempotent_request MUST be set if this is non-null */ grpc_metadata_batch *send_initial_metadata; @@ -252,6 +256,7 @@ void grpc_transport_stream_op_add_close(grpc_transport_stream_op *op, gpr_slice *optional_message); char *grpc_transport_stream_op_string(grpc_transport_stream_op *op); +char *grpc_transport_op_string(grpc_transport_op *op); /* Send a batch of operations on a transport @@ -293,6 +298,10 @@ char *grpc_transport_get_peer(grpc_exec_ctx *exec_ctx, /* Allocate a grpc_transport_op, and preconfigure the on_consumed closure to \a on_consumed and then delete the returned transport op */ grpc_transport_op *grpc_make_transport_op(grpc_closure *on_consumed); +/* Allocate a grpc_transport_stream_op, and preconfigure the on_consumed closure + to \a on_consumed and then delete the returned transport op */ +grpc_transport_stream_op *grpc_make_transport_stream_op( + grpc_closure *on_consumed); #ifdef __cplusplus } diff --git a/src/core/lib/transport/transport_op_string.c b/src/core/lib/transport/transport_op_string.c index 138591db2a..533ec52077 100644 --- a/src/core/lib/transport/transport_op_string.c +++ b/src/core/lib/transport/transport_op_string.c @@ -41,6 +41,7 @@ #include <grpc/support/string_util.h> #include <grpc/support/useful.h> #include "src/core/lib/support/string.h" +#include "src/core/lib/transport/connectivity_state.h" /* These routines are here to facilitate debugging - they produce string representations of various transport data structures */ @@ -72,56 +73,51 @@ static void put_metadata_list(gpr_strvec *b, grpc_metadata_batch md) { char *grpc_transport_stream_op_string(grpc_transport_stream_op *op) { char *tmp; char *out; - int first = 1; gpr_strvec b; gpr_strvec_init(&b); + gpr_strvec_add( + &b, gpr_strdup(op->covered_by_poller ? "[COVERED]" : "[UNCOVERED]")); + if (op->send_initial_metadata != NULL) { - if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); - first = 0; + gpr_strvec_add(&b, gpr_strdup(" ")); gpr_strvec_add(&b, gpr_strdup("SEND_INITIAL_METADATA{")); put_metadata_list(&b, *op->send_initial_metadata); gpr_strvec_add(&b, gpr_strdup("}")); } if (op->send_message != NULL) { - if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); - first = 0; + gpr_strvec_add(&b, gpr_strdup(" ")); gpr_asprintf(&tmp, "SEND_MESSAGE:flags=0x%08x:len=%d", op->send_message->flags, op->send_message->length); gpr_strvec_add(&b, tmp); } if (op->send_trailing_metadata != NULL) { - if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); - first = 0; + gpr_strvec_add(&b, gpr_strdup(" ")); gpr_strvec_add(&b, gpr_strdup("SEND_TRAILING_METADATA{")); put_metadata_list(&b, *op->send_trailing_metadata); gpr_strvec_add(&b, gpr_strdup("}")); } if (op->recv_initial_metadata != NULL) { - if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); - first = 0; + gpr_strvec_add(&b, gpr_strdup(" ")); gpr_strvec_add(&b, gpr_strdup("RECV_INITIAL_METADATA")); } if (op->recv_message != NULL) { - if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); - first = 0; + gpr_strvec_add(&b, gpr_strdup(" ")); gpr_strvec_add(&b, gpr_strdup("RECV_MESSAGE")); } if (op->recv_trailing_metadata != NULL) { - if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); - first = 0; + gpr_strvec_add(&b, gpr_strdup(" ")); gpr_strvec_add(&b, gpr_strdup("RECV_TRAILING_METADATA")); } if (op->cancel_error != GRPC_ERROR_NONE) { - if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); - first = 0; + gpr_strvec_add(&b, gpr_strdup(" ")); const char *msg = grpc_error_string(op->cancel_error); gpr_asprintf(&tmp, "CANCEL:%s", msg); grpc_error_free_string(msg); @@ -129,8 +125,7 @@ char *grpc_transport_stream_op_string(grpc_transport_stream_op *op) { } if (op->close_error != GRPC_ERROR_NONE) { - if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); - first = 0; + gpr_strvec_add(&b, gpr_strdup(" ")); const char *msg = grpc_error_string(op->close_error); gpr_asprintf(&tmp, "CLOSE:%s", msg); grpc_error_free_string(msg); @@ -143,6 +138,82 @@ char *grpc_transport_stream_op_string(grpc_transport_stream_op *op) { return out; } +char *grpc_transport_op_string(grpc_transport_op *op) { + char *tmp; + char *out; + bool first = true; + + gpr_strvec b; + gpr_strvec_init(&b); + + if (op->on_connectivity_state_change != NULL) { + if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); + first = false; + if (op->connectivity_state != NULL) { + gpr_asprintf(&tmp, "ON_CONNECTIVITY_STATE_CHANGE:p=%p:from=%s", + op->on_connectivity_state_change, + grpc_connectivity_state_name(*op->connectivity_state)); + gpr_strvec_add(&b, tmp); + } else { + gpr_asprintf(&tmp, "ON_CONNECTIVITY_STATE_CHANGE:p=%p:unsubscribe", + op->on_connectivity_state_change); + gpr_strvec_add(&b, tmp); + } + } + + if (op->disconnect_with_error != GRPC_ERROR_NONE) { + if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); + first = false; + const char *err = grpc_error_string(op->disconnect_with_error); + gpr_asprintf(&tmp, "DISCONNECT:%s", err); + gpr_strvec_add(&b, tmp); + grpc_error_free_string(err); + } + + if (op->send_goaway) { + if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); + first = false; + char *msg = op->goaway_message == NULL + ? "null" + : gpr_dump_slice(*op->goaway_message, + GPR_DUMP_ASCII | GPR_DUMP_HEX); + gpr_asprintf(&tmp, "SEND_GOAWAY:status=%d:msg=%s", op->goaway_status, msg); + if (op->goaway_message != NULL) gpr_free(msg); + gpr_strvec_add(&b, tmp); + } + + if (op->set_accept_stream) { + if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); + first = false; + gpr_asprintf(&tmp, "SET_ACCEPT_STREAM:%p(%p,...)", op->set_accept_stream_fn, + op->set_accept_stream_user_data); + gpr_strvec_add(&b, tmp); + } + + if (op->bind_pollset != NULL) { + if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); + first = false; + gpr_strvec_add(&b, gpr_strdup("BIND_POLLSET")); + } + + if (op->bind_pollset_set != NULL) { + if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); + first = false; + gpr_strvec_add(&b, gpr_strdup("BIND_POLLSET_SET")); + } + + if (op->send_ping != NULL) { + if (!first) gpr_strvec_add(&b, gpr_strdup(" ")); + first = false; + gpr_strvec_add(&b, gpr_strdup("SEND_PING")); + } + + out = gpr_strvec_flatten(&b, NULL); + gpr_strvec_destroy(&b); + + return out; +} + void grpc_call_log_op(char *file, int line, gpr_log_severity severity, grpc_call_element *elem, grpc_transport_stream_op *op) { char *str = grpc_transport_stream_op_string(op); |