Replace the temporary std::vector copy with in-place compaction using a
read/write pointer pattern. This avoids a heap allocation+deallocation
cycle during scheduler cleanup, reducing heap fragmentation on
long-running ESP devices.
The new approach compacts valid items forward in the existing vector,
recycles removed items as they are encountered, then resizes the vector
(no reallocation since size only shrinks). Same O(n) complexity, same
behavior, zero allocations.
Eliminate redundant xTaskGetCurrentTaskHandle() and pcTaskGetName()
calls on the hot path by resolving the thread name once in log_vprintf_
and passing it through as const char* to all downstream functions.
- Main task fast path passes nullptr (no task handle lookup needed)
- Non-main thread path resolves name once, passes to both ring buffer
and emergency console fallback
- Unify log_vprintf_non_main_thread_ to single signature across platforms
- Change send_message_thread_safe() on all platforms from TaskHandle_t
to const char* thread_name
- Add TaskHandle_t overload for get_thread_name_ as primary on
ESP32/LibreTiny, with no-arg convenience wrapper
- Use std::span<char> for Host/Zephyr get_thread_name_ buffer parameter
- Document Zephyr single-task path thread safety limitation
When a subprocess exited, _proc_on_exit sent the exit event but never
closed the server-side WebSocket. This left zombie connections open
until the client eventually disconnected.
Read all available bytes in batches via read_array() instead of
byte-at-a-time read() calls. Each read() internally chains through
read_byte -> read_array(1) -> check_read_timeout_ -> available(),
resulting in 3 UART calls per byte. Batching reduces this
significantly.
Read all available bytes in batches via read_array() instead of
byte-at-a-time read() calls. Each read() internally chains through
read_byte -> read_array(1) -> check_read_timeout_ -> available(),
resulting in 3 UART calls per byte. At 256000 baud with ~235 bytes
per loop iteration, this was ~706 UART operations per loop call.
Batching reduces this to ~12.
Measured 33% reduction in loop time (2348ms -> 1577ms per 60s).
Add _to() variants that write into a stack buffer and return StringRef,
avoiding heap allocation on every publish. Update publish_state() in
climate (11 allocations eliminated), fan (3), cover (2), and valve (1)
to use the new stack-based getters.
The allocating getters are retained for setup-time paths (subscribe,
send_discovery, dump_config) where the allocation is one-time.
The on_log callback was constructing a temporary MQTTMessage struct which
copied the log_message_ topic string and created a new std::string from
the message buffer on every log line. This is called frequently at runtime
and contributes to heap fragmentation.
Use the const char* publish overload directly to avoid both allocations.
