Optimize Application class memory layout and reduce loop_interval size

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

esphome/components/api/api_connection.cpp

@@ -33,9 +33,14 @@ namespace api {
 // Since each message could contain multiple protobuf messages when using packet batching,
 // this limits the number of messages processed, not the number of TCP packets.
 static constexpr uint8_t MAX_MESSAGES_PER_LOOP = 5;
+static constexpr uint8_t MAX_PING_RETRIES = 60;
+static constexpr uint16_t PING_RETRY_INTERVAL = 1000;
+static constexpr uint32_t KEEPALIVE_DISCONNECT_TIMEOUT = (KEEPALIVE_TIMEOUT_MS * 5) / 2;
 
 static const char *const TAG = "api.connection";
+#ifdef USE_ESP32_CAMERA
 static const int ESP32_CAMERA_STOP_STREAM = 5000;
+#endif
 
 APIConnection::APIConnection(std::unique_ptr<socket::Socket> sock, APIServer *parent)
     : parent_(parent), initial_state_iterator_(this), list_entities_iterator_(this) {
@@ -60,10 +65,6 @@ uint32_t APIConnection::get_batch_delay_ms_() const { return this->parent_->get_
 void APIConnection::start() {
   this->last_traffic_ = App.get_loop_component_start_time();
 
-  // Set next_ping_retry_ to prevent immediate ping
-  // This ensures the first ping happens after the keepalive period
-  this->next_ping_retry_ = this->last_traffic_ + KEEPALIVE_TIMEOUT_MS;
-
   APIError err = this->helper_->init();
   if (err != APIError::OK) {
     on_fatal_error();
@@ -90,16 +91,6 @@ APIConnection::~APIConnection() {
 }
 
 void APIConnection::loop() {
-  if (this->remove_)
-    return;
-
-  if (!network::is_connected()) {
-    // when network is disconnected force disconnect immediately
-    // don't wait for timeout
-    this->on_fatal_error();
-    ESP_LOGW(TAG, "%s: Network unavailable; disconnecting", this->get_client_combined_info().c_str());
-    return;
-  }
   if (this->next_close_) {
     // requested a disconnect
     this->helper_->close();
@@ -152,39 +143,31 @@ void APIConnection::loop() {
 
   // Process deferred batch if scheduled
   if (this->deferred_batch_.batch_scheduled &&
-      App.get_loop_component_start_time() - this->deferred_batch_.batch_start_time >= this->get_batch_delay_ms_()) {
+      now - this->deferred_batch_.batch_start_time >= this->get_batch_delay_ms_()) {
     this->process_batch_();
   }
 
-  if (!this->list_entities_iterator_.completed())
+  if (!this->list_entities_iterator_.completed()) {
     this->list_entities_iterator_.advance();
-  if (!this->initial_state_iterator_.completed() && this->list_entities_iterator_.completed())
+  } else if (!this->initial_state_iterator_.completed()) {
     this->initial_state_iterator_.advance();
+  }
 
-  static uint8_t max_ping_retries = 60;
-  static uint16_t ping_retry_interval = 1000;
   if (this->sent_ping_) {
     // Disconnect if not responded within 2.5*keepalive
-    if (now - this->last_traffic_ > (KEEPALIVE_TIMEOUT_MS * 5) / 2) {
+    if (now - this->last_traffic_ > KEEPALIVE_DISCONNECT_TIMEOUT) {
       on_fatal_error();
       ESP_LOGW(TAG, "%s is unresponsive; disconnecting", this->get_client_combined_info().c_str());
     }
-  } else if (now - this->last_traffic_ > KEEPALIVE_TIMEOUT_MS && now > this->next_ping_retry_) {
+  } else if (now - this->last_traffic_ > KEEPALIVE_TIMEOUT_MS) {
     ESP_LOGVV(TAG, "Sending keepalive PING");
     this->sent_ping_ = this->send_message(PingRequest());
     if (!this->sent_ping_) {
-      this->next_ping_retry_ = now + ping_retry_interval;
-      this->ping_retries_++;
-      std::string warn_str = str_sprintf("%s: Sending keepalive failed %u time(s);",
-                                         this->get_client_combined_info().c_str(), this->ping_retries_);
-      if (this->ping_retries_ >= max_ping_retries) {
-        on_fatal_error();
-        ESP_LOGE(TAG, "%s disconnecting", warn_str.c_str());
-      } else if (this->ping_retries_ >= 10) {
-        ESP_LOGW(TAG, "%s retrying in %u ms", warn_str.c_str(), ping_retry_interval);
-      } else {
-        ESP_LOGD(TAG, "%s retrying in %u ms", warn_str.c_str(), ping_retry_interval);
-      }
+      // If we can't send the ping request directly (tx_buffer full),
+      // schedule it at the front of the batch so it will be sent with priority
+      ESP_LOGVV(TAG, "Failed to send ping directly, scheduling at front of batch");
+      this->schedule_message_front_(nullptr, &APIConnection::try_send_ping_request, PingRequest::MESSAGE_TYPE);
+      this->sent_ping_ = true;  // Mark as sent to avoid scheduling multiple pings
     }
   }
 
@@ -207,22 +190,20 @@ void APIConnection::loop() {
     // bool done = 3;
     buffer.encode_bool(3, done);
 
-    bool success = this->send_buffer(buffer, 44);
+    bool success = this->send_buffer(buffer, CameraImageResponse::MESSAGE_TYPE);
 
     if (success) {
       this->image_reader_.consume_data(to_send);
-    }
-    if (success && done) {
-      this->image_reader_.return_image();
+      if (done) {
+        this->image_reader_.return_image();
+      }
     }
   }
 #endif
 
-  if (state_subs_at_ != -1) {
+  if (state_subs_at_ >= 0) {
     const auto &subs = this->parent_->get_state_subs();
-    if (state_subs_at_ >= (int) subs.size()) {
-      state_subs_at_ = -1;
-    } else {
+    if (state_subs_at_ < static_cast<int>(subs.size())) {
       auto &it = subs[state_subs_at_];
       SubscribeHomeAssistantStateResponse resp;
       resp.entity_id = it.entity_id;
@@ -231,6 +212,8 @@ void APIConnection::loop() {
       if (this->send_message(resp)) {
         state_subs_at_++;
       }
+    } else {
+      state_subs_at_ = -1;
     }
   }
 }
@@ -1440,7 +1423,7 @@ void APIConnection::alarm_control_panel_command(const AlarmControlPanelCommandRe
 
 #ifdef USE_EVENT
 void APIConnection::send_event(event::Event *event, const std::string &event_type) {
-  this->schedule_message_(event, MessageCreator(event_type, EventResponse::MESSAGE_TYPE), EventResponse::MESSAGE_TYPE);
+  this->schedule_message_(event, MessageCreator(event_type), EventResponse::MESSAGE_TYPE);
 }
 void APIConnection::send_event_info(event::Event *event) {
   this->schedule_message_(event, &APIConnection::try_send_event_info, ListEntitiesEventResponse::MESSAGE_TYPE);
@@ -1760,6 +1743,11 @@ void APIConnection::DeferredBatch::add_item(EntityBase *entity, MessageCreator c
   items.emplace_back(entity, std::move(creator), message_type);
 }
 
+void APIConnection::DeferredBatch::add_item_front(EntityBase *entity, MessageCreator creator, uint16_t message_type) {
+  // Insert at front for high priority messages (no deduplication check)
+  items.insert(items.begin(), BatchItem(entity, std::move(creator), message_type));
+}
+
 bool APIConnection::schedule_batch_() {
   if (!this->deferred_batch_.batch_scheduled) {
     this->deferred_batch_.batch_scheduled = true;
@@ -1795,7 +1783,8 @@ void APIConnection::process_batch_() {
     const auto &item = this->deferred_batch_.items[0];
 
     // Let the creator calculate size and encode if it fits
-    uint16_t payload_size = item.creator(item.entity, this, std::numeric_limits<uint16_t>::max(), true);
+    uint16_t payload_size =
+        item.creator(item.entity, this, std::numeric_limits<uint16_t>::max(), true, item.message_type);
 
     if (payload_size > 0 &&
         this->send_buffer(ProtoWriteBuffer{&this->parent_->get_shared_buffer_ref()}, item.message_type)) {
@@ -1845,7 +1834,7 @@ void APIConnection::process_batch_() {
   for (const auto &item : this->deferred_batch_.items) {
     // Try to encode message
     // The creator will calculate overhead to determine if the message fits
-    uint16_t payload_size = item.creator(item.entity, this, remaining_size, false);
+    uint16_t payload_size = item.creator(item.entity, this, remaining_size, false, item.message_type);
 
     if (payload_size == 0) {
       // Message won't fit, stop processing
@@ -1908,21 +1897,23 @@ void APIConnection::process_batch_() {
 }
 
 uint16_t APIConnection::MessageCreator::operator()(EntityBase *entity, APIConnection *conn, uint32_t remaining_size,
-                                                   bool is_single) const {
-  switch (message_type_) {
-    case 0:  // Function pointer
-      return data_.ptr(entity, conn, remaining_size, is_single);
-
+                                                   bool is_single, uint16_t message_type) const {
+  if (has_tagged_string_ptr_()) {
+    // Handle string-based messages
+    switch (message_type) {
 #ifdef USE_EVENT
-    case EventResponse::MESSAGE_TYPE: {
-      auto *e = static_cast<event::Event *>(entity);
-      return APIConnection::try_send_event_response(e, *data_.string_ptr, conn, remaining_size, is_single);
-    }
+      case EventResponse::MESSAGE_TYPE: {
+        auto *e = static_cast<event::Event *>(entity);
+        return APIConnection::try_send_event_response(e, *get_string_ptr_(), conn, remaining_size, is_single);
+      }
 #endif
-
-    default:
-      // Should not happen, return 0 to indicate no message
-      return 0;
+      default:
+        // Should not happen, return 0 to indicate no message
+        return 0;
+    }
+  } else {
+    // Function pointer case
+    return data_.ptr(entity, conn, remaining_size, is_single);
   }
 }
 
@@ -1938,6 +1929,12 @@ uint16_t APIConnection::try_send_disconnect_request(EntityBase *entity, APIConne
   return encode_message_to_buffer(req, DisconnectRequest::MESSAGE_TYPE, conn, remaining_size, is_single);
 }
 
+uint16_t APIConnection::try_send_ping_request(EntityBase *entity, APIConnection *conn, uint32_t remaining_size,
+                                              bool is_single) {
+  PingRequest req;
+  return encode_message_to_buffer(req, PingRequest::MESSAGE_TYPE, conn, remaining_size, is_single);
+}
+
 uint16_t APIConnection::get_estimated_message_size(uint16_t message_type) {
   // Use generated ESTIMATED_SIZE constants from each message type
   switch (message_type) {

esphome/components/api/api_connection.h

@@ -185,7 +185,6 @@ class APIConnection : public APIServerConnection {
   void on_disconnect_response(const DisconnectResponse &value) override;
   void on_ping_response(const PingResponse &value) override {
     // we initiated ping
-    this->ping_retries_ = 0;
     this->sent_ping_ = false;
   }
   void on_home_assistant_state_response(const HomeAssistantStateResponse &msg) override;
@@ -441,13 +440,16 @@ class APIConnection : public APIServerConnection {
   // Helper function to get estimated message size for buffer pre-allocation
   static uint16_t get_estimated_message_size(uint16_t message_type);
 
+  // Batch message method for ping requests
+  static uint16_t try_send_ping_request(EntityBase *entity, APIConnection *conn, uint32_t remaining_size,
+                                        bool is_single);
+
   // Pointers first (4 bytes each, naturally aligned)
   std::unique_ptr<APIFrameHelper> helper_;
   APIServer *parent_;
 
   // 4-byte aligned types
   uint32_t last_traffic_;
-  uint32_t next_ping_retry_{0};
   int state_subs_at_ = -1;
 
   // Strings (12 bytes each on 32-bit)
@@ -470,8 +472,7 @@ class APIConnection : public APIServerConnection {
   bool sent_ping_{false};
   bool service_call_subscription_{false};
   bool next_close_ = false;
-  uint8_t ping_retries_{0};
-  // 8 bytes used, no padding needed
+  // 7 bytes used, 1 byte padding
 
   // Larger objects at the end
   InitialStateIterator initial_state_iterator_;
@@ -483,55 +484,57 @@ class APIConnection : public APIServerConnection {
   // Function pointer type for message encoding
   using MessageCreatorPtr = uint16_t (*)(EntityBase *, APIConnection *, uint32_t remaining_size, bool is_single);
 
-  // Optimized MessageCreator class using union dispatch
+  // Optimized MessageCreator class using tagged pointer
   class MessageCreator {
+    // Ensure pointer alignment allows LSB tagging
+    static_assert(alignof(std::string *) > 1, "String pointer alignment must be > 1 for LSB tagging");
+
    public:
-    // Constructor for function pointer (message_type = 0)
-    MessageCreator(MessageCreatorPtr ptr) : message_type_(0) { data_.ptr = ptr; }
+    // Constructor for function pointer
+    MessageCreator(MessageCreatorPtr ptr) {
+      // Function pointers are always aligned, so LSB is 0
+      data_.ptr = ptr;
+    }
 
     // Constructor for string state capture
-    MessageCreator(const std::string &value, uint16_t msg_type) : message_type_(msg_type) {
-      data_.string_ptr = new std::string(value);
+    explicit MessageCreator(const std::string &str_value) {
+      // Allocate string and tag the pointer
+      auto *str = new std::string(str_value);
+      // Set LSB to 1 to indicate string pointer
+      data_.tagged = reinterpret_cast<uintptr_t>(str) | 1;
     }
 
     // Destructor
     ~MessageCreator() {
-      // Clean up string data for string-based message types
-      if (uses_string_data_()) {
-        delete data_.string_ptr;
+      if (has_tagged_string_ptr_()) {
+        delete get_string_ptr_();
       }
     }
 
     // Copy constructor
-    MessageCreator(const MessageCreator &other) : message_type_(other.message_type_) {
-      if (message_type_ == 0) {
-        data_.ptr = other.data_.ptr;
-      } else if (uses_string_data_()) {
-        data_.string_ptr = new std::string(*other.data_.string_ptr);
+    MessageCreator(const MessageCreator &other) {
+      if (other.has_tagged_string_ptr_()) {
+        auto *str = new std::string(*other.get_string_ptr_());
+        data_.tagged = reinterpret_cast<uintptr_t>(str) | 1;
       } else {
-        data_ = other.data_;  // For POD types
+        data_ = other.data_;
       }
     }
 
     // Move constructor
-    MessageCreator(MessageCreator &&other) noexcept : data_(other.data_), message_type_(other.message_type_) {
-      other.message_type_ = 0;  // Reset other to function pointer type
-      other.data_.ptr = nullptr;
-    }
+    MessageCreator(MessageCreator &&other) noexcept : data_(other.data_) { other.data_.ptr = nullptr; }
 
     // Assignment operators (needed for batch deduplication)
     MessageCreator &operator=(const MessageCreator &other) {
       if (this != &other) {
         // Clean up current string data if needed
-        if (uses_string_data_()) {
-          delete data_.string_ptr;
+        if (has_tagged_string_ptr_()) {
+          delete get_string_ptr_();
         }
         // Copy new data
-        message_type_ = other.message_type_;
-        if (other.message_type_ == 0) {
-          data_.ptr = other.data_.ptr;
-        } else if (other.uses_string_data_()) {
-          data_.string_ptr = new std::string(*other.data_.string_ptr);
+        if (other.has_tagged_string_ptr_()) {
+          auto *str = new std::string(*other.get_string_ptr_());
+          data_.tagged = reinterpret_cast<uintptr_t>(str) | 1;
         } else {
           data_ = other.data_;
         }
@@ -542,30 +545,35 @@ class APIConnection : public APIServerConnection {
     MessageCreator &operator=(MessageCreator &&other) noexcept {
       if (this != &other) {
         // Clean up current string data if needed
-        if (uses_string_data_()) {
-          delete data_.string_ptr;
+        if (has_tagged_string_ptr_()) {
+          delete get_string_ptr_();
         }
         // Move data
-        message_type_ = other.message_type_;
         data_ = other.data_;
         // Reset other to safe state
-        other.message_type_ = 0;
         other.data_.ptr = nullptr;
       }
       return *this;
     }
 
-    // Call operator
-    uint16_t operator()(EntityBase *entity, APIConnection *conn, uint32_t remaining_size, bool is_single) const;
+    // Call operator - now accepts message_type as parameter
+    uint16_t operator()(EntityBase *entity, APIConnection *conn, uint32_t remaining_size, bool is_single,
+                        uint16_t message_type) const;
 
    private:
-    // Helper to check if this message type uses heap-allocated strings
-    bool uses_string_data_() const { return message_type_ == EventResponse::MESSAGE_TYPE; }
-    union CreatorData {
-      MessageCreatorPtr ptr;    // 8 bytes
-      std::string *string_ptr;  // 8 bytes
-    } data_;                    // 8 bytes
-    uint16_t message_type_;     // 2 bytes (0 = function ptr, >0 = state capture)
+    // Check if this contains a string pointer
+    bool has_tagged_string_ptr_() const { return (data_.tagged & 1) != 0; }
+
+    // Get the actual string pointer (clears the tag bit)
+    std::string *get_string_ptr_() const {
+      // NOLINTNEXTLINE(performance-no-int-to-ptr)
+      return reinterpret_cast<std::string *>(data_.tagged & ~uintptr_t(1));
+    }
+
+    union {
+      MessageCreatorPtr ptr;
+      uintptr_t tagged;
+    } data_;  // 4 bytes on 32-bit
   };
 
   // Generic batching mechanism for both state updates and entity info
@@ -591,6 +599,8 @@ class APIConnection : public APIServerConnection {
 
     // Add item to the batch
     void add_item(EntityBase *entity, MessageCreator creator, uint16_t message_type);
+    // Add item to the front of the batch (for high priority messages like ping)
+    void add_item_front(EntityBase *entity, MessageCreator creator, uint16_t message_type);
     void clear() {
       items.clear();
       batch_scheduled = false;
@@ -630,6 +640,12 @@ class APIConnection : public APIServerConnection {
   bool schedule_message_(EntityBase *entity, MessageCreatorPtr function_ptr, uint16_t message_type) {
     return schedule_message_(entity, MessageCreator(function_ptr), message_type);
   }
+
+  // Helper function to schedule a high priority message at the front of the batch
+  bool schedule_message_front_(EntityBase *entity, MessageCreatorPtr function_ptr, uint16_t message_type) {
+    this->deferred_batch_.add_item_front(entity, MessageCreator(function_ptr), message_type);
+    return this->schedule_batch_();
+  }
 };
 
 }  // namespace api

esphome/components/api/api_frame_helper.cpp

@@ -66,6 +66,17 @@ const char *api_error_to_str(APIError err) {
   return "UNKNOWN";
 }
 
+// Default implementation for loop - handles sending buffered data
+APIError APIFrameHelper::loop() {
+  if (!this->tx_buf_.empty()) {
+    APIError err = try_send_tx_buf_();
+    if (err != APIError::OK && err != APIError::WOULD_BLOCK) {
+      return err;
+    }
+  }
+  return APIError::OK;  // Convert WOULD_BLOCK to OK to avoid connection termination
+}
+
 // Helper method to buffer data from IOVs
 void APIFrameHelper::buffer_data_from_iov_(const struct iovec *iov, int iovcnt, uint16_t total_write_len) {
   SendBuffer buffer;
@@ -287,13 +298,8 @@ APIError APINoiseFrameHelper::loop() {
     }
   }
 
-  if (!this->tx_buf_.empty()) {
-    APIError err = try_send_tx_buf_();
-    if (err != APIError::OK && err != APIError::WOULD_BLOCK) {
-      return err;
-    }
-  }
-  return APIError::OK;  // Convert WOULD_BLOCK to OK to avoid connection termination
+  // Use base class implementation for buffer sending
+  return APIFrameHelper::loop();
 }
 
 /** Read a packet into the rx_buf_. If successful, stores frame data in the frame parameter
@@ -339,17 +345,15 @@ APIError APINoiseFrameHelper::try_read_frame_(ParsedFrame *frame) {
       return APIError::WOULD_BLOCK;
     }
 
+    if (rx_header_buf_[0] != 0x01) {
+      state_ = State::FAILED;
+      HELPER_LOG("Bad indicator byte %u", rx_header_buf_[0]);
+      return APIError::BAD_INDICATOR;
+    }
     // header reading done
   }
 
   // read body
-  uint8_t indicator = rx_header_buf_[0];
-  if (indicator != 0x01) {
-    state_ = State::FAILED;
-    HELPER_LOG("Bad indicator byte %u", indicator);
-    return APIError::BAD_INDICATOR;
-  }
-
   uint16_t msg_size = (((uint16_t) rx_header_buf_[1]) << 8) | rx_header_buf_[2];
 
   if (state_ != State::DATA && msg_size > 128) {
@@ -595,10 +599,6 @@ APIError APINoiseFrameHelper::read_packet(ReadPacketBuffer *buffer) {
     return APIError::BAD_DATA_PACKET;
   }
 
-  // uint16_t type;
-  // uint16_t data_len;
-  // uint8_t *data;
-  // uint8_t *padding;  zero or more bytes to fill up the rest of the packet
   uint16_t type = (((uint16_t) msg_data[0]) << 8) | msg_data[1];
   uint16_t data_len = (((uint16_t) msg_data[2]) << 8) | msg_data[3];
   if (data_len > msg_size - 4) {
@@ -831,18 +831,12 @@ APIError APIPlaintextFrameHelper::init() {
   state_ = State::DATA;
   return APIError::OK;
 }
-/// Not used for plaintext
 APIError APIPlaintextFrameHelper::loop() {
   if (state_ != State::DATA) {
     return APIError::BAD_STATE;
   }
-  if (!this->tx_buf_.empty()) {
-    APIError err = try_send_tx_buf_();
-    if (err != APIError::OK && err != APIError::WOULD_BLOCK) {
-      return err;
-    }
-  }
-  return APIError::OK;  // Convert WOULD_BLOCK to OK to avoid connection termination
+  // Use base class implementation for buffer sending
+  return APIFrameHelper::loop();
 }
 
 /** Read a packet into the rx_buf_. If successful, stores frame data in the frame parameter

esphome/components/api/api_frame_helper.h

@@ -38,7 +38,7 @@ struct PacketInfo {
       : message_type(type), offset(off), payload_size(size), padding(0) {}
 };
 
-enum class APIError : int {
+enum class APIError : uint16_t {
   OK = 0,
   WOULD_BLOCK = 1001,
   BAD_HANDSHAKE_PACKET_LEN = 1002,
@@ -74,7 +74,7 @@ class APIFrameHelper {
   }
   virtual ~APIFrameHelper() = default;
   virtual APIError init() = 0;
-  virtual APIError loop() = 0;
+  virtual APIError loop();
   virtual APIError read_packet(ReadPacketBuffer *buffer) = 0;
   bool can_write_without_blocking() { return state_ == State::DATA && tx_buf_.empty(); }
   std::string getpeername() { return socket_->getpeername(); }

esphome/components/api/api_server.cpp

@@ -47,6 +47,11 @@ void APIServer::setup() {
   }
 #endif
 
+  // Schedule reboot if no clients connect within timeout
+  if (this->reboot_timeout_ != 0) {
+    this->schedule_reboot_timeout_();
+  }
+
   this->socket_ = socket::socket_ip_loop_monitored(SOCK_STREAM, 0);  // monitored for incoming connections
   if (this->socket_ == nullptr) {
     ESP_LOGW(TAG, "Could not create socket");
@@ -106,8 +111,6 @@ void APIServer::setup() {
   }
 #endif
 
-  this->last_connected_ = App.get_loop_component_start_time();
-
 #ifdef USE_ESP32_CAMERA
   if (esp32_camera::global_esp32_camera != nullptr && !esp32_camera::global_esp32_camera->is_internal()) {
     esp32_camera::global_esp32_camera->add_image_callback(
@@ -121,6 +124,16 @@ void APIServer::setup() {
 #endif
 }
 
+void APIServer::schedule_reboot_timeout_() {
+  this->status_set_warning();
+  this->set_timeout("api_reboot", this->reboot_timeout_, []() {
+    if (!global_api_server->is_connected()) {
+      ESP_LOGE(TAG, "No clients; rebooting");
+      App.reboot();
+    }
+  });
+}
+
 void APIServer::loop() {
   // Accept new clients only if the socket exists and has incoming connections
   if (this->socket_ && this->socket_->ready()) {
@@ -130,51 +143,61 @@ void APIServer::loop() {
       auto sock = this->socket_->accept_loop_monitored((struct sockaddr *) &source_addr, &addr_len);
       if (!sock)
         break;
-      ESP_LOGD(TAG, "Accepted %s", sock->getpeername().c_str());
+      ESP_LOGD(TAG, "Accept %s", sock->getpeername().c_str());
 
       auto *conn = new APIConnection(std::move(sock), this);
       this->clients_.emplace_back(conn);
       conn->start();
+
+      // Clear warning status and cancel reboot when first client connects
+      if (this->clients_.size() == 1 && this->reboot_timeout_ != 0) {
+        this->status_clear_warning();
+        this->cancel_timeout("api_reboot");
+      }
     }
   }
 
+  if (this->clients_.empty()) {
+    return;
+  }
+
   // Process clients and remove disconnected ones in a single pass
-  if (!this->clients_.empty()) {
-    size_t client_index = 0;
-    while (client_index < this->clients_.size()) {
-      auto &client = this->clients_[client_index];
-
-      if (client->remove_) {
-        // Handle disconnection
-        this->client_disconnected_trigger_->trigger(client->client_info_, client->client_peername_);
-        ESP_LOGV(TAG, "Removing connection to %s", client->client_info_.c_str());
-
-        // Swap with the last element and pop (avoids expensive vector shifts)
-        if (client_index < this->clients_.size() - 1) {
-          std::swap(this->clients_[client_index], this->clients_.back());
-        }
-        this->clients_.pop_back();
-        // Don't increment client_index since we need to process the swapped element
-      } else {
-        // Process active client
-        client->loop();
-        client_index++;  // Move to next client
-      }
+  // Check network connectivity once for all clients
+  if (!network::is_connected()) {
+    // Network is down - disconnect all clients
+    for (auto &client : this->clients_) {
+      client->on_fatal_error();
+      ESP_LOGW(TAG, "%s: Network down; disconnect", client->get_client_combined_info().c_str());
     }
+    // Continue to process and clean up the clients below
   }
 
-  if (this->reboot_timeout_ != 0) {
-    const uint32_t now = App.get_loop_component_start_time();
-    if (!this->is_connected()) {
-      if (now - this->last_connected_ > this->reboot_timeout_) {
-        ESP_LOGE(TAG, "No client connected; rebooting");
-        App.reboot();
-      }
-      this->status_set_warning();
-    } else {
-      this->last_connected_ = now;
-      this->status_clear_warning();
+  size_t client_index = 0;
+  while (client_index < this->clients_.size()) {
+    auto &client = this->clients_[client_index];
+
+    if (!client->remove_) {
+      // Common case: process active client
+      client->loop();
+      client_index++;
+      continue;
     }
+
+    // Rare case: handle disconnection
+    this->client_disconnected_trigger_->trigger(client->client_info_, client->client_peername_);
+    ESP_LOGV(TAG, "Remove connection %s", client->client_info_.c_str());
+
+    // Swap with the last element and pop (avoids expensive vector shifts)
+    if (client_index < this->clients_.size() - 1) {
+      std::swap(this->clients_[client_index], this->clients_.back());
+    }
+    this->clients_.pop_back();
+
+    // Schedule reboot when last client disconnects
+    if (this->clients_.empty() && this->reboot_timeout_ != 0) {
+      this->schedule_reboot_timeout_();
+    }
+    // Don't increment client_index since we need to process the swapped element
   }
 }
 
@@ -503,8 +526,8 @@ void APIServer::on_shutdown() {
   for (auto &c : this->clients_) {
     if (!c->send_message(DisconnectRequest())) {
       // If we can't send the disconnect request directly (tx_buffer full),
-      // schedule it in the batch so it will be sent with the 5ms timer
-      c->schedule_message_(nullptr, &APIConnection::try_send_disconnect_request, DisconnectRequest::MESSAGE_TYPE);
+      // schedule it at the front of the batch so it will be sent with priority
+      c->schedule_message_front_(nullptr, &APIConnection::try_send_disconnect_request, DisconnectRequest::MESSAGE_TYPE);
     }
   }
 }

esphome/components/api/api_server.h

@@ -142,6 +142,7 @@ class APIServer : public Component, public Controller {
   }
 
  protected:
+  void schedule_reboot_timeout_();
   // Pointers and pointer-like types first (4 bytes each)
   std::unique_ptr<socket::Socket> socket_ = nullptr;
   Trigger<std::string, std::string> *client_connected_trigger_ = new Trigger<std::string, std::string>();
@@ -150,7 +151,6 @@ class APIServer : public Component, public Controller {
   // 4-byte aligned types
   uint32_t reboot_timeout_{300000};
   uint32_t batch_delay_{100};
-  uint32_t last_connected_{0};
 
   // Vectors and strings (12 bytes each on 32-bit)
   std::vector<std::unique_ptr<APIConnection>> clients_;

esphome/components/gpio/binary_sensor/__init__.py

@@ -10,11 +10,24 @@ GPIOBinarySensor = gpio_ns.class_(
     "GPIOBinarySensor", binary_sensor.BinarySensor, cg.Component
 )
 
+CONF_USE_INTERRUPT = "use_interrupt"
+CONF_INTERRUPT_TYPE = "interrupt_type"
+
+INTERRUPT_TYPES = {
+    "RISING": gpio_ns.INTERRUPT_RISING_EDGE,
+    "FALLING": gpio_ns.INTERRUPT_FALLING_EDGE,
+    "ANY": gpio_ns.INTERRUPT_ANY_EDGE,
+}
+
 CONFIG_SCHEMA = (
     binary_sensor.binary_sensor_schema(GPIOBinarySensor)
     .extend(
         {
             cv.Required(CONF_PIN): pins.gpio_input_pin_schema,
+            cv.Optional(CONF_USE_INTERRUPT, default=True): cv.boolean,
+            cv.Optional(CONF_INTERRUPT_TYPE, default="ANY"): cv.enum(
+                INTERRUPT_TYPES, upper=True
+            ),
         }
     )
     .extend(cv.COMPONENT_SCHEMA)
@@ -27,3 +40,7 @@ async def to_code(config):
 
     pin = await cg.gpio_pin_expression(config[CONF_PIN])
     cg.add(var.set_pin(pin))
+
+    cg.add(var.set_use_interrupt(config[CONF_USE_INTERRUPT]))
+    if config[CONF_USE_INTERRUPT]:
+        cg.add(var.set_interrupt_type(INTERRUPT_TYPES[config[CONF_INTERRUPT_TYPE]]))

esphome/components/gpio/binary_sensor/gpio_binary_sensor.cpp

@@ -6,17 +6,91 @@ namespace gpio {
 
 static const char *const TAG = "gpio.binary_sensor";
 
+void IRAM_ATTR GPIOBinarySensorStore::gpio_intr(GPIOBinarySensorStore *arg) {
+  bool new_state = arg->isr_pin_.digital_read();
+  if (new_state != arg->last_state_) {
+    arg->state_ = new_state;
+    arg->last_state_ = new_state;
+    arg->changed_ = true;
+    // Wake up the component from its disabled loop state
+    if (arg->component_ != nullptr) {
+      arg->component_->enable_loop_soon_any_context();
+    }
+  }
+}
+
+void GPIOBinarySensorStore::setup(InternalGPIOPin *pin, gpio::InterruptType type, Component *component) {
+  pin->setup();
+  this->isr_pin_ = pin->to_isr();
+  this->component_ = component;
+
+  // Read initial state
+  this->last_state_ = pin->digital_read();
+  this->state_ = this->last_state_;
+
+  // Attach interrupt - from this point on, any changes will be caught by the interrupt
+  pin->attach_interrupt(&GPIOBinarySensorStore::gpio_intr, this, type);
+}
+
 void GPIOBinarySensor::setup() {
-  this->pin_->setup();
-  this->publish_initial_state(this->pin_->digital_read());
+  if (this->use_interrupt_ && !this->pin_->is_internal()) {
+    ESP_LOGD(TAG, "GPIO is not internal, falling back to polling mode");
+    this->use_interrupt_ = false;
+  }
+
+  if (this->use_interrupt_) {
+    auto *internal_pin = static_cast<InternalGPIOPin *>(this->pin_);
+    this->store_.setup(internal_pin, this->interrupt_type_, this);
+    this->publish_initial_state(this->store_.get_state());
+  } else {
+    this->pin_->setup();
+    this->publish_initial_state(this->pin_->digital_read());
+  }
 }
 
 void GPIOBinarySensor::dump_config() {
   LOG_BINARY_SENSOR("", "GPIO Binary Sensor", this);
   LOG_PIN("  Pin: ", this->pin_);
+  const char *mode = this->use_interrupt_ ? "interrupt" : "polling";
+  ESP_LOGCONFIG(TAG, "  Mode: %s", mode);
+  if (this->use_interrupt_) {
+    const char *interrupt_type;
+    switch (this->interrupt_type_) {
+      case gpio::INTERRUPT_RISING_EDGE:
+        interrupt_type = "RISING_EDGE";
+        break;
+      case gpio::INTERRUPT_FALLING_EDGE:
+        interrupt_type = "FALLING_EDGE";
+        break;
+      case gpio::INTERRUPT_ANY_EDGE:
+        interrupt_type = "ANY_EDGE";
+        break;
+      default:
+        interrupt_type = "UNKNOWN";
+        break;
+    }
+    ESP_LOGCONFIG(TAG, "  Interrupt Type: %s", interrupt_type);
+  }
 }
 
-void GPIOBinarySensor::loop() { this->publish_state(this->pin_->digital_read()); }
+void GPIOBinarySensor::loop() {
+  if (this->use_interrupt_) {
+    if (this->store_.is_changed()) {
+      // Clear the flag immediately to minimize the window where we might miss changes
+      this->store_.clear_changed();
+      // Read the state and publish it
+      // Note: If the ISR fires between clear_changed() and get_state(), that's fine -
+      // we'll process the new change on the next loop iteration
+      bool state = this->store_.get_state();
+      this->publish_state(state);
+    } else {
+      // No changes, disable the loop until the next interrupt
+      this->disable_loop();
+    }
+  } else {
+    this->publish_state(this->pin_->digital_read());
+  }
+}
 
 float GPIOBinarySensor::get_setup_priority() const { return setup_priority::HARDWARE; }
 

esphome/components/gpio/binary_sensor/gpio_binary_sensor.h

@@ -2,14 +2,51 @@
 
 #include "esphome/core/component.h"
 #include "esphome/core/hal.h"
+#include "esphome/core/helpers.h"
 #include "esphome/components/binary_sensor/binary_sensor.h"
 
 namespace esphome {
 namespace gpio {
 
+// Store class for ISR data (no vtables, ISR-safe)
+class GPIOBinarySensorStore {
+ public:
+  void setup(InternalGPIOPin *pin, gpio::InterruptType type, Component *component);
+
+  static void gpio_intr(GPIOBinarySensorStore *arg);
+
+  bool get_state() const {
+    // No lock needed: state_ is atomically updated by ISR
+    // Volatile ensures we read the latest value
+    return this->state_;
+  }
+
+  bool is_changed() const {
+    // Simple read of volatile bool - no clearing here
+    return this->changed_;
+  }
+
+  void clear_changed() {
+    // Separate method to clear the flag
+    this->changed_ = false;
+  }
+
+ protected:
+  ISRInternalGPIOPin isr_pin_;
+  volatile bool state_{false};
+  volatile bool last_state_{false};
+  volatile bool changed_{false};
+  Component *component_{nullptr};  // Pointer to the component for enable_loop_soon_any_context()
+};
+
 class GPIOBinarySensor : public binary_sensor::BinarySensor, public Component {
  public:
+  // No destructor needed: ESPHome components are created at boot and live forever.
+  // Interrupts are only detached on reboot when memory is cleared anyway.
+
   void set_pin(GPIOPin *pin) { pin_ = pin; }
+  void set_use_interrupt(bool use_interrupt) { use_interrupt_ = use_interrupt; }
+  void set_interrupt_type(gpio::InterruptType type) { interrupt_type_ = type; }
   // ========== INTERNAL METHODS ==========
   // (In most use cases you won't need these)
   /// Setup pin
@@ -22,6 +59,9 @@ class GPIOBinarySensor : public binary_sensor::BinarySensor, public Component {
 
  protected:
   GPIOPin *pin_;
+  bool use_interrupt_{true};
+  gpio::InterruptType interrupt_type_{gpio::INTERRUPT_ANY_EDGE};
+  GPIOBinarySensorStore store_;
 };
 
 }  // namespace gpio

esphome/components/logger/logger.cpp

@@ -48,6 +48,11 @@ void HOT Logger::log_vprintf_(uint8_t level, const char *tag, int line, const ch
   // For non-main tasks, queue the message for callbacks - but only if we have any callbacks registered
   message_sent =
       this->log_buffer_->send_message_thread_safe(level, tag, static_cast<uint16_t>(line), current_task, format, args);
+  if (message_sent) {
+    // Enable logger loop to process the buffered message
+    // This is safe to call from any context including ISRs
+    this->enable_loop_soon_any_context();
+  }
 #endif  // USE_ESPHOME_TASK_LOG_BUFFER
 
   // Emergency console logging for non-main tasks when ring buffer is full or disabled
@@ -139,6 +144,10 @@ Logger::Logger(uint32_t baud_rate, size_t tx_buffer_size) : baud_rate_(baud_rate
 #ifdef USE_ESPHOME_TASK_LOG_BUFFER
 void Logger::init_log_buffer(size_t total_buffer_size) {
   this->log_buffer_ = esphome::make_unique<logger::TaskLogBuffer>(total_buffer_size);
+
+  // Start with loop disabled when using task buffer (unless using USB CDC)
+  // The loop will be enabled automatically when messages arrive
+  this->disable_loop_when_buffer_empty_();
 }
 #endif
 
@@ -189,6 +198,10 @@ void Logger::loop() {
         this->write_msg_(this->tx_buffer_);
       }
     }
+  } else {
+    // No messages to process, disable loop if appropriate
+    // This reduces overhead when there's no async logging activity
+    this->disable_loop_when_buffer_empty_();
   }
 #endif
 }

esphome/components/logger/logger.h

@@ -358,6 +358,26 @@ class Logger : public Component {
     static const uint16_t RESET_COLOR_LEN = strlen(ESPHOME_LOG_RESET_COLOR);
     this->write_body_to_buffer_(ESPHOME_LOG_RESET_COLOR, RESET_COLOR_LEN, buffer, buffer_at, buffer_size);
   }
+
+#ifdef USE_ESP32
+  // Disable loop when task buffer is empty (with USB CDC check)
+  inline void disable_loop_when_buffer_empty_() {
+    // Thread safety note: This is safe even if another task calls enable_loop_soon_any_context()
+    // concurrently. If that happens between our check and disable_loop(), the enable request
+    // will be processed on the next main loop iteration since:
+    // - disable_loop() takes effect immediately
+    // - enable_loop_soon_any_context() sets a pending flag that's checked at loop start
+#if defined(USE_LOGGER_USB_CDC) && defined(USE_ARDUINO)
+    // Only disable if not using USB CDC (which needs loop for connection detection)
+    if (this->uart_ != UART_SELECTION_USB_CDC) {
+      this->disable_loop();
+    }
+#else
+    // No USB CDC support, always safe to disable
+    this->disable_loop();
+#endif
+  }
+#endif
 };
 extern Logger *global_logger;  // NOLINT(cppcoreguidelines-avoid-non-const-global-variables)
 

esphome/components/runtime_stats/__init__.py

@@ -0,0 +1,26 @@
+"""
+Runtime statistics component for ESPHome.
+"""
+
+import esphome.codegen as cg
+import esphome.config_validation as cv
+
+DEPENDENCIES = []
+
+CONF_ENABLED = "enabled"
+CONF_LOG_INTERVAL = "log_interval"
+
+CONFIG_SCHEMA = cv.Schema(
+    {
+        cv.Optional(CONF_ENABLED, default=True): cv.boolean,
+        cv.Optional(
+            CONF_LOG_INTERVAL, default=60000
+        ): cv.positive_time_period_milliseconds,
+    }
+)
+
+
+async def to_code(config):
+    """Generate code for the runtime statistics component."""
+    cg.add(cg.App.set_runtime_stats_enabled(config[CONF_ENABLED]))
+    cg.add(cg.App.set_runtime_stats_log_interval(config[CONF_LOG_INTERVAL]))

esphome/core/application.cpp

@@ -136,6 +136,10 @@ void Application::loop() {
   this->in_loop_ = false;
   this->app_state_ = new_app_state;
 
+  // Process any pending runtime stats printing after all components have run
+  // This ensures stats printing doesn't affect component timing measurements
+  runtime_stats.process_pending_stats(last_op_end_time);
+
   // Use the last component's end time instead of calling millis() again
   auto elapsed = last_op_end_time - this->last_loop_;
   if (elapsed >= this->loop_interval_ || HighFrequencyLoopRequester::is_high_frequency()) {

esphome/core/application.h

@@ -1,5 +1,7 @@
 #pragma once
 
+#include <algorithm>
+#include <limits>
 #include <string>
 #include <vector>
 #include "esphome/core/component.h"
@@ -7,6 +9,7 @@
 #include "esphome/core/hal.h"
 #include "esphome/core/helpers.h"
 #include "esphome/core/preferences.h"
+#include "esphome/core/runtime_stats.h"
 #include "esphome/core/scheduler.h"
 
 #ifdef USE_DEVICES
@@ -337,9 +340,23 @@ class Application {
    *
    * @param loop_interval The interval in milliseconds to run the core loop at. Defaults to 16 milliseconds.
    */
-  void set_loop_interval(uint32_t loop_interval) { this->loop_interval_ = loop_interval; }
+  void set_loop_interval(uint32_t loop_interval) {
+    this->loop_interval_ = std::min(loop_interval, static_cast<uint32_t>(std::numeric_limits<uint16_t>::max()));
+  }
 
-  uint32_t get_loop_interval() const { return this->loop_interval_; }
+  uint32_t get_loop_interval() const { return static_cast<uint32_t>(this->loop_interval_); }
+
+  /** Enable or disable runtime statistics collection.
+   *
+   * @param enable Whether to enable runtime statistics collection.
+   */
+  void set_runtime_stats_enabled(bool enable) { runtime_stats.set_enabled(enable); }
+
+  /** Set the interval at which runtime statistics are logged.
+   *
+   * @param interval The interval in milliseconds between logging of runtime statistics.
+   */
+  void set_runtime_stats_log_interval(uint32_t interval) { runtime_stats.set_log_interval(interval); }
 
   void schedule_dump_config() { this->dump_config_at_ = 0; }
 
@@ -618,6 +635,17 @@ class Application {
   /// Perform a delay while also monitoring socket file descriptors for readiness
   void yield_with_select_(uint32_t delay_ms);
 
+  // === Member variables ordered by size to minimize padding ===
+
+  // Pointer-sized members first
+  Component *current_component_{nullptr};
+  const char *comment_{nullptr};
+  const char *compilation_time_{nullptr};
+
+  // size_t members
+  size_t dump_config_at_{SIZE_MAX};
+
+  // Vectors (largest members)
   std::vector<Component *> components_{};
 
   // Partitioned vector design for looping components
@@ -637,11 +665,6 @@ class Application {
   //   and active_end_ is incremented
   // - This eliminates branch mispredictions from flag checking in the hot loop
   std::vector<Component *> looping_components_{};
-  uint16_t looping_components_active_end_{0};
-
-  // For safe reentrant modifications during iteration
-  uint16_t current_loop_index_{0};
-  bool in_loop_{false};
 
 #ifdef USE_DEVICES
   std::vector<Device *> devices_{};
@@ -713,26 +736,39 @@ class Application {
   std::vector<update::UpdateEntity *> updates_{};
 #endif
 
+#ifdef USE_SOCKET_SELECT_SUPPORT
+  std::vector<int> socket_fds_;  // Vector of all monitored socket file descriptors
+#endif
+
+  // String members
   std::string name_;
   std::string friendly_name_;
-  const char *comment_{nullptr};
-  const char *compilation_time_{nullptr};
-  bool name_add_mac_suffix_;
+
+  // 4-byte members
   uint32_t last_loop_{0};
-  uint32_t loop_interval_{16};
-  size_t dump_config_at_{SIZE_MAX};
-  uint8_t app_state_{0};
-  volatile bool has_pending_enable_loop_requests_{false};
-  Component *current_component_{nullptr};
   uint32_t loop_component_start_time_{0};
 
 #ifdef USE_SOCKET_SELECT_SUPPORT
-  // Socket select management
-  std::vector<int> socket_fds_;     // Vector of all monitored socket file descriptors
+  int max_fd_{-1};  // Highest file descriptor number for select()
+#endif
+
+  // 2-byte members (grouped together for alignment)
+  uint16_t loop_interval_{16};  // Loop interval in ms (max 65535ms = 65.5 seconds)
+  uint16_t looping_components_active_end_{0};
+  uint16_t current_loop_index_{0};  // For safe reentrant modifications during iteration
+
+  // 1-byte members (grouped together to minimize padding)
+  uint8_t app_state_{0};
+  bool name_add_mac_suffix_;
+  bool in_loop_{false};
+  volatile bool has_pending_enable_loop_requests_{false};
+
+#ifdef USE_SOCKET_SELECT_SUPPORT
   bool socket_fds_changed_{false};  // Flag to rebuild base_read_fds_ when socket_fds_ changes
-  int max_fd_{-1};                  // Highest file descriptor number for select()
-  fd_set base_read_fds_{};          // Cached fd_set rebuilt only when socket_fds_ changes
-  fd_set read_fds_{};               // Working fd_set for select(), copied from base_read_fds_
+
+  // Variable-sized members at end
+  fd_set base_read_fds_{};  // Cached fd_set rebuilt only when socket_fds_ changes
+  fd_set read_fds_{};       // Working fd_set for select(), copied from base_read_fds_
 #endif
 };
 

esphome/core/automation.h

@@ -27,20 +27,67 @@ template<typename T, typename... X> class TemplatableValue {
  public:
   TemplatableValue() : type_(NONE) {}
 
-  template<typename F, enable_if_t<!is_invocable<F, X...>::value, int> = 0>
-  TemplatableValue(F value) : type_(VALUE), value_(std::move(value)) {}
+  template<typename F, enable_if_t<!is_invocable<F, X...>::value, int> = 0> TemplatableValue(F value) : type_(VALUE) {
+    new (&this->value_) T(std::move(value));
+  }
 
-  template<typename F, enable_if_t<is_invocable<F, X...>::value, int> = 0>
-  TemplatableValue(F f) : type_(LAMBDA), f_(f) {}
+  template<typename F, enable_if_t<is_invocable<F, X...>::value, int> = 0> TemplatableValue(F f) : type_(LAMBDA) {
+    this->f_ = new std::function<T(X...)>(std::move(f));
+  }
+
+  // Copy constructor
+  TemplatableValue(const TemplatableValue &other) : type_(other.type_) {
+    if (type_ == VALUE) {
+      new (&this->value_) T(other.value_);
+    } else if (type_ == LAMBDA) {
+      this->f_ = new std::function<T(X...)>(*other.f_);
+    }
+  }
+
+  // Move constructor
+  TemplatableValue(TemplatableValue &&other) noexcept : type_(other.type_) {
+    if (type_ == VALUE) {
+      new (&this->value_) T(std::move(other.value_));
+    } else if (type_ == LAMBDA) {
+      this->f_ = other.f_;
+      other.f_ = nullptr;
+    }
+    other.type_ = NONE;
+  }
+
+  // Assignment operators
+  TemplatableValue &operator=(const TemplatableValue &other) {
+    if (this != &other) {
+      this->~TemplatableValue();
+      new (this) TemplatableValue(other);
+    }
+    return *this;
+  }
+
+  TemplatableValue &operator=(TemplatableValue &&other) noexcept {
+    if (this != &other) {
+      this->~TemplatableValue();
+      new (this) TemplatableValue(std::move(other));
+    }
+    return *this;
+  }
+
+  ~TemplatableValue() {
+    if (type_ == VALUE) {
+      this->value_.~T();
+    } else if (type_ == LAMBDA) {
+      delete this->f_;
+    }
+  }
 
   bool has_value() { return this->type_ != NONE; }
 
   T value(X... x) {
     if (this->type_ == LAMBDA) {
-      return this->f_(x...);
+      return (*this->f_)(x...);
     }
     // return value also when none
-    return this->value_;
+    return this->type_ == VALUE ? this->value_ : T{};
   }
 
   optional<T> optional_value(X... x) {
@@ -58,14 +105,16 @@ template<typename T, typename... X> class TemplatableValue {
   }
 
  protected:
-  enum {
+  enum : uint8_t {
     NONE,
     VALUE,
     LAMBDA,
   } type_;
 
-  T value_{};
-  std::function<T(X...)> f_{};
+  union {
+    T value_;
+    std::function<T(X...)> *f_;
+  };
 };
 
 /** Base class for all automation conditions.

esphome/core/component.cpp

@@ -303,6 +303,9 @@ uint32_t WarnIfComponentBlockingGuard::finish() {
   uint32_t curr_time = millis();
 
   uint32_t blocking_time = curr_time - this->started_;
+
+  // Record component runtime stats
+  runtime_stats.record_component_time(this->component_, blocking_time, curr_time);
   bool should_warn;
   if (this->component_ != nullptr) {
     should_warn = this->component_->should_warn_of_blocking(blocking_time);

esphome/core/component.h

@@ -6,6 +6,7 @@
 #include <string>
 
 #include "esphome/core/optional.h"
+#include "esphome/core/runtime_stats.h"
 
 namespace esphome {
 

esphome/core/component_iterator.cpp

@@ -375,7 +375,7 @@ void ComponentIterator::advance() {
   }
 
   if (advance_platform) {
-    this->state_ = static_cast<IteratorState>(static_cast<uint32_t>(this->state_) + 1);
+    this->state_ = static_cast<IteratorState>(static_cast<uint16_t>(this->state_) + 1);
     this->at_ = 0;
   } else if (success) {
     this->at_++;

esphome/core/component_iterator.h

@@ -93,7 +93,9 @@ class ComponentIterator {
   virtual bool on_end();
 
  protected:
-  enum class IteratorState {
+  // Iterates over all ESPHome entities (sensors, switches, lights, etc.)
+  // Supports up to 256 entity types and up to 65,535 entities of each type
+  enum class IteratorState : uint8_t {
     NONE = 0,
     BEGIN,
 #ifdef USE_BINARY_SENSOR
@@ -167,7 +169,7 @@ class ComponentIterator {
 #endif
     MAX,
   } state_{IteratorState::NONE};
-  size_t at_{0};
+  uint16_t at_{0};  // Supports up to 65,535 entities per type
   bool include_internal_{false};
 };
 

esphome/core/defines.h

@@ -132,6 +132,8 @@
 
 // ESP32-specific feature flags
 #ifdef USE_ESP32
+#define USE_ESPHOME_TASK_LOG_BUFFER
+
 #define USE_BLUETOOTH_PROXY
 #define USE_CAPTIVE_PORTAL
 #define USE_ESP32_BLE

esphome/core/runtime_stats.cpp

@@ -0,0 +1,92 @@
+#include "esphome/core/runtime_stats.h"
+#include "esphome/core/component.h"
+#include <algorithm>
+
+namespace esphome {
+
+RuntimeStatsCollector runtime_stats;
+
+void RuntimeStatsCollector::record_component_time(Component *component, uint32_t duration_ms, uint32_t current_time) {
+  if (!this->enabled_ || component == nullptr)
+    return;
+
+  // Check if we have cached the name for this component
+  auto name_it = this->component_names_cache_.find(component);
+  if (name_it == this->component_names_cache_.end()) {
+    // First time seeing this component, cache its name
+    const char *source = component->get_component_source();
+    this->component_names_cache_[component] = source;
+    this->component_stats_[source].record_time(duration_ms);
+  } else {
+    // Use cached name - no string operations, just map lookup
+    this->component_stats_[name_it->second].record_time(duration_ms);
+  }
+
+  // If next_log_time_ is 0, initialize it
+  if (this->next_log_time_ == 0) {
+    this->next_log_time_ = current_time + this->log_interval_;
+    return;
+  }
+
+  // Don't print stats here anymore - let process_pending_stats handle it
+}
+
+void RuntimeStatsCollector::log_stats_() {
+  ESP_LOGI(RUNTIME_TAG, "Component Runtime Statistics");
+  ESP_LOGI(RUNTIME_TAG, "Period stats (last %" PRIu32 "ms):", this->log_interval_);
+
+  // First collect stats we want to display
+  std::vector<ComponentStatPair> stats_to_display;
+
+  for (const auto &it : this->component_stats_) {
+    const ComponentRuntimeStats &stats = it.second;
+    if (stats.get_period_count() > 0) {
+      ComponentStatPair pair = {it.first, &stats};
+      stats_to_display.push_back(pair);
+    }
+  }
+
+  // Sort by period runtime (descending)
+  std::sort(stats_to_display.begin(), stats_to_display.end(), std::greater<ComponentStatPair>());
+
+  // Log top components by period runtime
+  for (const auto &it : stats_to_display) {
+    const std::string &source = it.name;
+    const ComponentRuntimeStats *stats = it.stats;
+
+    ESP_LOGI(RUNTIME_TAG, "  %s: count=%" PRIu32 ", avg=%.2fms, max=%" PRIu32 "ms, total=%" PRIu32 "ms", source.c_str(),
+             stats->get_period_count(), stats->get_period_avg_time_ms(), stats->get_period_max_time_ms(),
+             stats->get_period_time_ms());
+  }
+
+  // Log total stats since boot
+  ESP_LOGI(RUNTIME_TAG, "Total stats (since boot):");
+
+  // Re-sort by total runtime for all-time stats
+  std::sort(stats_to_display.begin(), stats_to_display.end(),
+            [](const ComponentStatPair &a, const ComponentStatPair &b) {
+              return a.stats->get_total_time_ms() > b.stats->get_total_time_ms();
+            });
+
+  for (const auto &it : stats_to_display) {
+    const std::string &source = it.name;
+    const ComponentRuntimeStats *stats = it.stats;
+
+    ESP_LOGI(RUNTIME_TAG, "  %s: count=%" PRIu32 ", avg=%.2fms, max=%" PRIu32 "ms, total=%" PRIu32 "ms", source.c_str(),
+             stats->get_total_count(), stats->get_total_avg_time_ms(), stats->get_total_max_time_ms(),
+             stats->get_total_time_ms());
+  }
+}
+
+void RuntimeStatsCollector::process_pending_stats(uint32_t current_time) {
+  if (!this->enabled_ || this->next_log_time_ == 0)
+    return;
+
+  if (current_time >= this->next_log_time_) {
+    this->log_stats_();
+    this->reset_stats_();
+    this->next_log_time_ = current_time + this->log_interval_;
+  }
+}
+
+}  // namespace esphome

esphome/core/runtime_stats.h

@@ -0,0 +1,121 @@
+#pragma once
+
+#include <map>
+#include <string>
+#include <vector>
+#include <cstdint>
+#include <algorithm>
+#include "esphome/core/helpers.h"
+#include "esphome/core/log.h"
+
+namespace esphome {
+
+static const char *const RUNTIME_TAG = "runtime";
+
+class Component;  // Forward declaration
+
+class ComponentRuntimeStats {
+ public:
+  ComponentRuntimeStats()
+      : period_count_(0),
+        total_count_(0),
+        period_time_ms_(0),
+        total_time_ms_(0),
+        period_max_time_ms_(0),
+        total_max_time_ms_(0) {}
+
+  void record_time(uint32_t duration_ms) {
+    // Update period counters
+    this->period_count_++;
+    this->period_time_ms_ += duration_ms;
+    if (duration_ms > this->period_max_time_ms_)
+      this->period_max_time_ms_ = duration_ms;
+
+    // Update total counters
+    this->total_count_++;
+    this->total_time_ms_ += duration_ms;
+    if (duration_ms > this->total_max_time_ms_)
+      this->total_max_time_ms_ = duration_ms;
+  }
+
+  void reset_period_stats() {
+    this->period_count_ = 0;
+    this->period_time_ms_ = 0;
+    this->period_max_time_ms_ = 0;
+  }
+
+  // Period stats (reset each logging interval)
+  uint32_t get_period_count() const { return this->period_count_; }
+  uint32_t get_period_time_ms() const { return this->period_time_ms_; }
+  uint32_t get_period_max_time_ms() const { return this->period_max_time_ms_; }
+  float get_period_avg_time_ms() const {
+    return this->period_count_ > 0 ? this->period_time_ms_ / static_cast<float>(this->period_count_) : 0.0f;
+  }
+
+  // Total stats (persistent until reboot)
+  uint32_t get_total_count() const { return this->total_count_; }
+  uint32_t get_total_time_ms() const { return this->total_time_ms_; }
+  uint32_t get_total_max_time_ms() const { return this->total_max_time_ms_; }
+  float get_total_avg_time_ms() const {
+    return this->total_count_ > 0 ? this->total_time_ms_ / static_cast<float>(this->total_count_) : 0.0f;
+  }
+
+ protected:
+  // Period stats (reset each logging interval)
+  uint32_t period_count_;
+  uint32_t period_time_ms_;
+  uint32_t period_max_time_ms_;
+
+  // Total stats (persistent until reboot)
+  uint32_t total_count_;
+  uint32_t total_time_ms_;
+  uint32_t total_max_time_ms_;
+};
+
+// For sorting components by run time
+struct ComponentStatPair {
+  std::string name;
+  const ComponentRuntimeStats *stats;
+
+  bool operator>(const ComponentStatPair &other) const {
+    // Sort by period time as that's what we're displaying in the logs
+    return stats->get_period_time_ms() > other.stats->get_period_time_ms();
+  }
+};
+
+class RuntimeStatsCollector {
+ public:
+  RuntimeStatsCollector() : log_interval_(60000), next_log_time_(0), enabled_(true) {}
+
+  void set_log_interval(uint32_t log_interval) { this->log_interval_ = log_interval; }
+  uint32_t get_log_interval() const { return this->log_interval_; }
+
+  void set_enabled(bool enabled) { this->enabled_ = enabled; }
+  bool is_enabled() const { return this->enabled_; }
+
+  void record_component_time(Component *component, uint32_t duration_ms, uint32_t current_time);
+
+  // Process any pending stats printing (should be called after component loop)
+  void process_pending_stats(uint32_t current_time);
+
+ protected:
+  void log_stats_();
+
+  void reset_stats_() {
+    for (auto &it : this->component_stats_) {
+      it.second.reset_period_stats();
+    }
+  }
+
+  // Back to string keys, but we'll cache the source name per component
+  std::map<std::string, ComponentRuntimeStats> component_stats_;
+  std::map<Component *, std::string> component_names_cache_;
+  uint32_t log_interval_;
+  uint32_t next_log_time_;
+  bool enabled_;
+};
+
+// Global instance for runtime stats collection
+extern RuntimeStatsCollector runtime_stats;
+
+}  // namespace esphome

tests/integration/fixtures/api_reboot_timeout.yaml

@@ -0,0 +1,7 @@
+esphome:
+  name: api-reboot-test
+host:
+api:
+  reboot_timeout: 0.5s  # Very short timeout for fast testing
+logger:
+  level: DEBUG

tests/integration/test_api_reboot_timeout.py

@@ -0,0 +1,35 @@
+"""Test API server reboot timeout functionality."""
+
+import asyncio
+import re
+
+import pytest
+
+from .types import RunCompiledFunction
+
+
+@pytest.mark.asyncio
+async def test_api_reboot_timeout(
+    yaml_config: str,
+    run_compiled: RunCompiledFunction,
+) -> None:
+    """Test that the device reboots when no API clients connect within the timeout."""
+    loop = asyncio.get_running_loop()
+    reboot_future = loop.create_future()
+    reboot_pattern = re.compile(r"No clients; rebooting")
+
+    def check_output(line: str) -> None:
+        """Check output for reboot message."""
+        if not reboot_future.done() and reboot_pattern.search(line):
+            reboot_future.set_result(True)
+
+    # Run the device without connecting any API client
+    async with run_compiled(yaml_config, line_callback=check_output):
+        # Wait for reboot with timeout
+        # (0.5s reboot timeout + some margin for processing)
+        try:
+            await asyncio.wait_for(reboot_future, timeout=2.0)
+        except asyncio.TimeoutError:
+            pytest.fail("Device did not reboot within expected timeout")
+
+    # Test passes if we get here - reboot was detected