- _get_level_hashes: use separate SQL without prefix64<=? for level 0.
UINT64_MAX cast to sqlite3_int64 becomes -1, making prefix64<=-1 always
false for positive prefixes → bitmap always 0 → sync never exchanges data.
- member_sync_init: add _rebuild_all_trees to scan all peers_* tables and
call merkle_sync_recompute_path for each member. Ensures merkle_tree_hash
has entries for members added before merkle sync existed.
- Add 2-byte dlen prefix to terminal BATCH/HASHES data so receiver knows
exact byte count — fixes rem=0 bug that truncated multi-item BATCHes.
- Fix synthetic HASHES prefix_bytes: use merkle_sync_prefix_bytes(next_lvl)
for both declaration and _prefix_write, not current level's pb_i.
- Fix initiator never getting _session_done: responder sends confirmation
HASHES(is_data=1,level=0) before _session_done so initiator's done_cb fires.
- Add level==0 guard in _member_get_items/_member_update_bucket_hash
to prevent SQL queries with bogus mask on virtual root level.
- _get_level_hashes: remove (0x1F << parent_shift) from range_end so
bitmap only includes children of the actual parent prefix (not all
parents). Add level 0 virtual root that queries all level-1 entries.
- merkle_sync_start: send HASHES at level 0 (virtual root) instead of
level 1/prefix 0. Root-level REQUEST prefix reconstruction now uses
the correct level-1 prefix from DB.
- _handle_batch: use merkle_sync_prefix_bytes(next_lvl) for synthetic
child HASHES instead of current level's pb_i (could be wrong size).
- Remove timeout/retry mechanism: delete _session_timeout_cb,
_session_start_timer, timer/retries fields. ETCP provides reliable
transport — no need for application-layer retries.
_prefix_write was writing the least significant pb bytes of the 64-bit
prefix, but the protocol expects the most significant pb bytes (bits
63..64-pb*8). For pb<=4 these ranges are disjoint → wire always got
zeros. _prefix_read also lacked the final left-shift to align the
result at bit 63.
This caused all level>=2 merkle sync requests/batches to receive
prefix=0 instead of the real prefix, making get_items/update_bucket_hash
find zero members → handle_batch len=7 with no data → sync completed
without actually exchanging member data.
После cascade_from в db_handle_send_data оповещаем ВСЕХ пиров
с synced_pos >= fix_from о том что цепочка изменилась — откатываем
их synced_pos. Это устраняет неэффективность конкурентных sync:
раньше только пир-отправитель обновлялся, остальные видели устаревшие
хеши и обнаруживали ложное расхождение на следующем цикле.
- si_send_data_batch: неправильный SQL SELECT с несуществующей колонкой author
вместо node_id → ломал всю передачу данных в REFINE/SEND_DATA
- Добавлена проверка si_prep и возврат int с кодом ошибки
- Все вызовы si_send_data_batch проверяют возврат, при ошибке сброс sync_state
- db_handle_push: добавлен db_cascade_from после вставки (раньше chain_hash
не пересчитывался после PUSH)
- db_handle_refine: DEBUG_WARN при scnt==0 (fm за пределами диапазона)
- Добавлена публичная db_sync_chain_verify() для верификации цепочек
- test_db_sync: фаза 4 (divergence через разныe hashes + REFINE/SEND_DATA),
фаза 5 (PUSH без хвоста + cascade verification)
- Убран мёртвый код db_sync_on_new_conn
- chatgui: адаптивная раскраска пузырей исходящих сообщений
SIGINT on Linux can be delivered to any thread (not just worker).
signal_handler uses g_currentNode->requestStop(); with thread_local,
g_currentNode was nullptr in non-worker threads — Ctrl-C did nothing.
Removing thread_local fixes signal delivery: g_currentNode is set once
before signal handler is registered, pointer read is atomic on x86/x64.
- AnimTimer::shutdown(): stop QTimer, clear m_icons before icon destruction
- ~MainWindow: AnimTimer::shutdown() -> cleanupAnimations() (correct order)
- main: replace ExitProcess with TerminateProcess(GetCurrentProcess,0)
TerminateProcess skips DLL_PROCESS_DETACH, avoiding libwinpthread
pthread_join hang on rlottie/miniaudio threads.
rlottie C API has no global shutdown - internal RleTaskScheduler threads
cannot be stopped through public API. Same for miniaudio WASAPI threads.
All owned resources already freed before TerminateProcess.
Root cause: librlottie RleTaskScheduler creates 32 worker threads on first
lottie_animation_from_data call. These pthreads wait in WaitForMultipleObjects(INFINITE)
and never exit. On MinGW, exit()/pthread_join blocks forever.
Changes:
- messagelist: s_icons moved to file-static; cleanupAnimations() clears them,
calling lottie_animation_destroy BEFORE static destruction
- mainwindow: ~MainWindow calls MessageList::cleanupAnimations() after node stop
- main: ShutdownSentinel counts threads at exit; if >4 remaining, calls
ExitProcess(0) to bypass CRT pthread_join hang
- main: if rlottie properly shuts down (<=4 threads), normal return (no ExitProcess)
Three diagnostic points covering the gap between ~MainWindow done and CRT exit:
- DbManager: destroyed — confirms local dtors started after ~MainWindow
- SoundManager: static destructor entered — confirms static destruction phase
- main: static destructors done (g_shutdown_sentinel dtor) — confirms ALL statics done
If process hangs, we know exactly which phase failed.
- Unified format: 'sync [SHORT_TBL:PEER_HI] <-/-> action: key_data'
- Each log line explains what the algorithm is doing and why
- Added batch progress tracking in dump-all and tail SEND_DATA loops
- Track dropped records (duplicates, bad_sigs, parse_fails) in SEND_DATA recv
- Show synced_pos transitions (old->new) in SEND_DATA and SYNC_DONE
- Show all launched syncs in PEER CHECK log
- Convert all silent returns to ERROR logs with [TBL:NODE] context
- Convert old WARN/ERROR formats to new sync [TBL:NODE] format
- Add sync_state reset details on ERROR and timeout paths
Signal handler via g_shutdown + uasync_wakeup ensures clean uasync poll exit.
updateReadPosition() now saves both lastReadMsgId and lastPosMsgId on every
scroll stop (300ms debounce), not just when read position changes.
connect(stopped, quit) triggers graceful Qt shutdown after uasync cleanup.
QString::fromUtf8(ch_id) read past length-prefixed string, capturing garbage.
Use fromUtf8(ch_id, ch_id_len) — length already available from bridge but was unused.
Messages now appear on refresh and unread badge correctly matches channel.
Replace gui_bridge_post_uasync_fn (async queue lost on shutdown) with direct
chat_core_save_ui_state (sync SQLite write). SQLite WAL handles concurrent access.
Add DEBUG_DEBUG traces in closeEvent, ~MainWindow, aboutToQuit.
Add author_node_id:8 to GUI_EVT_MSG_RECEIVED data format.
Callback now includes uint64_t author parameter.
SoundManager::play('msg_incoming') only when author != myNodeId.
- Same-channel refresh: save scroll pos in memory, restore from memory (no DB race)
- Force m_readDebounce->start() after loadChannel (scrollToBottom may not trigger valueChanged)
- DEBUG_DEBUG logs in updateReadPosition, saveChannelState, onMessageReceived, setChannelUnreadCount
- Debounce timer (300ms) advances lastReadMsgId only when message is fully in viewport
- Unread badge no longer resets on channel select — only when actual scroll reads msgs
- saveChannelState() uses tracked m_lastReadMsgId instead of viewport bottom
- getChannels() reads lastReadMsgId from ui_state (was reading unused channels column)
file(GLOB) in libutun/CMakeLists.txt picked up mplay.c, mdb_dump.c and other
standalone LMDB tools that are not library code and fail on Windows. Filter them
out, keep only mdb.c + midl.c.
Store read/scroll positions per channel via ui_state table (key-value).
Draw red +N badge on channel list items for unread messages.
Fix getMessages() to fetch last N messages (was fetching first N).
chat_channel.c: new chat_core_create_channel_auto(name) — generates
X25519/Ed25519 keys, random channel_id, Ed25519 signature, creates channel.
chat_core_create_channel_auto_trampoline for gui_bridge_post_uasync_fn.
mainwindow.cpp: onCreateGroupRequested() reduced from 100 to 15 lines.
Removed openssl/evp.h and openssl/rand.h includes from UI code.
Evgeny