[api] Peel first write iteration, inline socket writes, zero-gap batch encoding (#15063)

This commit is contained in:
J. Nick Koston
2026-04-08 11:05:53 -10:00
committed by GitHub
parent 4a18ef87d7
commit 576d89a82a
10 changed files with 374 additions and 253 deletions
+14 -50
View File
@@ -406,7 +406,7 @@ uint16_t APIConnection::fill_and_encode_entity_info(EntityBase *entity, InfoResp
#ifdef USE_DEVICES
msg.device_id = entity->get_device_id();
#endif
return encode_to_buffer(size_fn(&msg), encode_fn, &msg, conn, remaining_size);
return encode_to_buffer_slow(size_fn(&msg), encode_fn, &msg, conn, remaining_size);
}
uint16_t APIConnection::fill_and_encode_entity_info_with_device_class(EntityBase *entity, InfoResponseProtoMessage &msg,
@@ -2005,48 +2005,12 @@ bool APIConnection::send_message_(uint32_t payload_size, uint8_t message_type, M
encode_fn(msg, buffer PROTO_ENCODE_DEBUG_INIT(&shared_buf));
return this->send_buffer(ProtoWriteBuffer{&shared_buf}, message_type);
}
// Encodes a message to the buffer and returns the total number of bytes used,
// including header and footer overhead. Returns 0 if the message doesn't fit.
uint16_t APIConnection::encode_to_buffer(uint32_t calculated_size, MessageEncodeFn encode_fn, const void *msg,
APIConnection *conn, uint32_t remaining_size) {
#ifdef HAS_PROTO_MESSAGE_DUMP
if (conn->flags_.log_only_mode) {
auto *proto_msg = static_cast<const ProtoMessage *>(msg);
DumpBuffer dump_buf;
conn->log_send_message_(proto_msg->message_name(), proto_msg->dump_to(dump_buf));
return 1;
}
#endif
// Cache frame sizes to avoid repeated virtual calls
const uint8_t header_padding = conn->helper_->frame_header_padding();
const uint8_t footer_size = conn->helper_->frame_footer_size();
// encode_to_buffer is defined inline in api_connection.h (ESPHOME_ALWAYS_INLINE)
// Calculate total size with padding for buffer allocation
size_t total_calculated_size = calculated_size + header_padding + footer_size;
// Check if it fits
if (total_calculated_size > remaining_size)
return 0; // Doesn't fit
auto &shared_buf = conn->parent_->get_shared_buffer_ref();
size_t to_add;
if (conn->flags_.batch_first_message) {
// First message - buffer already prepared by caller, just clear flag
conn->flags_.batch_first_message = false;
to_add = calculated_size;
} else {
// Batch message second or later
// Reserve for full message, resize to include footer gap + header padding + payload
to_add = total_calculated_size;
}
shared_buf.resize(shared_buf.size() + to_add);
ProtoWriteBuffer buffer{&shared_buf, shared_buf.size() - calculated_size};
encode_fn(msg, buffer PROTO_ENCODE_DEBUG_INIT(&shared_buf));
// Return total size (header + payload + footer)
return static_cast<uint16_t>(total_calculated_size);
// Noinline version for cold paths — single shared copy
uint16_t APIConnection::encode_to_buffer_slow(uint32_t calculated_size, MessageEncodeFn encode_fn, const void *msg,
APIConnection *conn, uint32_t remaining_size) {
return encode_to_buffer(calculated_size, encode_fn, msg, conn, remaining_size);
}
bool APIConnection::send_buffer(ProtoWriteBuffer buffer, uint8_t message_type) {
const bool is_log_message = (message_type == SubscribeLogsResponse::MESSAGE_TYPE);
@@ -2173,17 +2137,15 @@ void APIConnection::process_batch_multi_(APIBuffer &shared_buf, size_t num_items
"MessageInfo must remain trivially destructible with this placement-new approach");
const size_t messages_to_process = std::min(num_items, MAX_MESSAGES_PER_BATCH);
const uint8_t frame_overhead = header_padding + footer_size;
// Stack-allocated array for message info
alignas(MessageInfo) char message_info_storage[MAX_MESSAGES_PER_BATCH * sizeof(MessageInfo)];
MessageInfo *message_info = reinterpret_cast<MessageInfo *>(message_info_storage);
size_t items_processed = 0;
uint16_t remaining_size = std::numeric_limits<uint16_t>::max();
// Track where each message's header padding begins in the buffer
// For plaintext: this is where the 6-byte header padding starts
// For noise: this is where the 7-byte header padding starts
// The actual message data follows after the header padding
// Track where each message's header begins in the buffer
// First message: offset 0 (max padding, may have unused leading bytes)
// Subsequent messages: offset points to exact header start (no gaps)
uint32_t current_offset = 0;
// Process items and encode directly to buffer (up to our limit)
@@ -2199,13 +2161,14 @@ void APIConnection::process_batch_multi_(APIBuffer &shared_buf, size_t num_items
}
// Message was encoded successfully
// payload_size is header_padding + actual payload size + footer_size
uint16_t proto_payload_size = payload_size - frame_overhead;
// payload_size = header_size + proto_payload_size + footer_size
uint16_t proto_payload_size = payload_size - this->batch_header_size_ - footer_size;
// Use placement new to construct MessageInfo in pre-allocated stack array
// This avoids default-constructing all MAX_MESSAGES_PER_BATCH elements
// Explicit destruction is not needed because MessageInfo is trivially destructible,
// as ensured by the static_assert in its definition.
new (&message_info[items_processed++]) MessageInfo(item.message_type, current_offset, proto_payload_size);
new (&message_info[items_processed++])
MessageInfo(item.message_type, current_offset, proto_payload_size, this->batch_header_size_);
// After first message, set remaining size to MAX_BATCH_PACKET_SIZE to avoid fragmentation
if (items_processed == 1) {
remaining_size = MAX_BATCH_PACKET_SIZE;
@@ -2255,6 +2218,7 @@ void APIConnection::process_batch_multi_(APIBuffer &shared_buf, size_t num_items
uint16_t APIConnection::dispatch_message_(const DeferredBatch::BatchItem &item, uint32_t remaining_size,
bool batch_first) {
this->flags_.batch_first_message = batch_first;
this->batch_message_type_ = item.message_type;
#ifdef USE_EVENT
// Events need aux_data_index to look up event type from entity
if (item.message_type == EventResponse::MESSAGE_TYPE) {