[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
+56 -8
View File
@@ -411,16 +411,59 @@ class APIConnection final : public APIServerConnectionBase {
// Non-template buffer management for send_message
bool send_message_(uint32_t payload_size, uint8_t message_type, MessageEncodeFn encode_fn, const void *msg);
// Non-template buffer management for batch encoding
static uint16_t encode_to_buffer(uint32_t calculated_size, MessageEncodeFn encode_fn, const void *msg,
APIConnection *conn, uint32_t remaining_size);
// Core batch encoding logic. Computes header size, checks fit, resizes buffer, encodes.
// ALWAYS_INLINE so the compiler can devirtualize encode_fn at hot call sites.
static inline uint16_t ESPHOME_ALWAYS_INLINE 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
const uint8_t footer_size = conn->helper_->frame_footer_size();
// Thin template wrapper — computes size, delegates buffer work to non-template helper
// First message uses max padding (already in buffer), subsequent use exact header size
size_t to_add;
if (conn->flags_.batch_first_message) {
conn->flags_.batch_first_message = false;
conn->batch_header_size_ = conn->helper_->frame_header_padding();
to_add = calculated_size;
} else {
conn->batch_header_size_ = conn->helper_->frame_header_size(calculated_size, conn->batch_message_type_);
to_add = calculated_size + conn->batch_header_size_ + footer_size;
}
// Check if it fits (using actual header size, not max padding)
uint16_t total_calculated_size = calculated_size + conn->batch_header_size_ + footer_size;
if (total_calculated_size > remaining_size)
return 0;
auto &shared_buf = conn->parent_->get_shared_buffer_ref();
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_calculated_size;
}
// Noinline version of encode_to_buffer for cold paths (entity info, zero-payload messages).
// All cold callers share this single copy instead of each getting an ALWAYS_INLINE expansion.
static uint16_t encode_to_buffer_slow(uint32_t calculated_size, MessageEncodeFn encode_fn, const void *msg,
APIConnection *conn, uint32_t remaining_size);
// Thin template wrapper — uses noinline encode_to_buffer_slow since
// encode_message_to_buffer callers are cold paths (zero-payload control messages).
// Hot paths (state/info) go through fill_and_encode_entity_state/info instead.
// batch_message_type_ is already set by dispatch_message_ before reaching here.
template<typename T> static uint16_t encode_message_to_buffer(T &msg, APIConnection *conn, uint32_t remaining_size) {
if constexpr (T::ESTIMATED_SIZE == 0) {
return encode_to_buffer(0, &encode_msg_noop, &msg, conn, remaining_size);
return encode_to_buffer_slow(0, &encode_msg_noop, &msg, conn, remaining_size);
} else {
return encode_to_buffer(msg.calculate_size(), &proto_encode_msg<T>, &msg, conn, remaining_size);
return encode_to_buffer_slow(msg.calculate_size(), &proto_encode_msg<T>, &msg, conn, remaining_size);
}
}
@@ -735,9 +778,14 @@ class APIConnection final : public APIServerConnectionBase {
// 2-byte types immediately after flags_ (no padding between them)
uint16_t client_api_version_major_{0};
uint16_t client_api_version_minor_{0};
// 1-byte type to fill padding
// 1-byte types to fill remaining space before next 4-byte boundary
ActiveIterator active_iterator_{ActiveIterator::NONE};
// Total: 2 (flags) + 2 + 2 + 1 = 7 bytes, then 1 byte padding to next 4-byte boundary
uint8_t batch_message_type_{0}; // Current message type during batch encoding
// Total: 2 (flags) + 2 + 2 + 1 + 1 = 8 bytes, aligned to 4-byte boundary
// Actual header size used by encode_to_buffer for the current message.
// Read by process_batch_multi_ to pass into MessageInfo.
uint8_t batch_header_size_{0};
uint32_t get_batch_delay_ms_() const { return this->parent_->get_batch_delay(); }
// Message will use 8 more bytes than the minimum size, and typical