Стресс-тест (Stage 7) ждал sync_pending==0, но счётчик был разбалансирован
(merkle_sync_start перетирает done_cb без вызова) и никогда не доходил до 0,
из-за чего тест всегда выжигал таймаут 20s.
Теперь сходимость определяется по реальному условию — совпадению корневых
хешей — финальными раундами синка, управляемыми таймерами и done-коллбэками.
Попутно: исправлен iter==50->25, удалён мёртвый код.
- etcp_link_enter_ready_tcp: убраны прямые записи got_initial_pkt/initialized/
links_up/tx_state и fire UP/INIT; переход через etcp_conn_ready +
loadbalancer_link_ready (как в UDP INIT-обработчиках)
- tcp_server_on_link: убран дублирующий gop/session-блок (проверка в
etcp_link_enter_ready_tcp единообразно), sc_init_ctx перед
sc_set_peer_public_key (фикс ctx not initialized)
- stcp_link_connect(etcp_link, addr, port) вместо config-структуры;
stcp_conn.etcp_conn для live-чтения gop/session_id при handshake
- новый модуль etcp_keepalive.c/h: keepalive-пакеты, таймер, адаптивный период,
согласование при handshake вынесены из etcp_connections.c как есть
- протокол KEEPALIVE_REQ/RESP (0x09/0x0A): мобильный узел при смене active mode
шлёт желаемый режим; отвечающий считает max (standby при любом standby), не-мобильный
применяет как есть; инициатор применяет согласованный режим после RESP
- режимы: NORMAL (3с, адаптивный рост), STANDBY (30с, ka_pinned=1 без роста)
- active mode как событие: utun_add/remove_activity_cbk; подписчики topo_group
(nodeinfo) и etcp_keepalive (рассылка REQ всем линкам)
- dedup double ETCP_CONN_STATUS_UP (UDP+TCP links) in topo_group_new_conn
- close conn_mgr handle on TIMEOUT in Phase2/3 (fix stuck 'already connecting')
- infinite Phase1→2→3 cycle with pause (1s active / 30s standby)
- CHAT_EVT_CONNECTING_NODES: yellow circle in member list while connecting
- Android: client_activity wiring via onStart/onStop + nativeSetActive
- function-level comments in topo_group_connect.c
- client_activity: desktop/server now ACTIVE by default, only mobile starts STANDBY
- topo_group_connect: endless Phase1→2→3→1 cycle until connected
- Pause between failed Phase2/3 attempts: 1s active, 30s standby
- on_down always triggers cycle_restart when active_conn_count=0
- re-init destroys previous state and starts fresh
- CHAT_EVT_CONNECTING_NODES event notifies GUI of current connecting nodes
- Android: ConfigProvider default mode='auto', PTT fix, UDP log timer fix
- android_jni_bridge: utun_bridge_set_debug_level actually calls debug_set_level / debug_set_category_level_by_name (was stub)
- SettingsScreen: CATEGORIES updated to match current g_debug_categories from debug_config.c
- ConfigProvider: persist category levels in [debug] section of generated INI config
- u_async: poll() log uses computed timeout_ms instead of confusing %d.%dms format
- u_async: timeout_node char name[16] → const char* name (no truncation, smaller struct)
- memory_pool: pool alloc/free debug logs moved from DEBUG to TRACE
- instance_lite: poll timeout changed from 10ms (100tb) to infinite (-1) — sleeps until traffic
- instance_lite: removed useless heartbeat timer
- etcp_link_enter_init: new link on live conn sends NOINIT, not full reset
- etcp_link_enter_reinit: don't conn_reinit on single link recovery, preserve inflight
- handle_nl_link: fall through to remove_iface_sockets when if_indextoname fails
- duplicate INIT: add INFO log for diagnostic
Test: sequence checking (gaps=0 dups=0), exact socket/link counts, stale link
detection, conn health checks, full disconnect+reconnect phase
When ip link del removes an interface, the kernel sends RTM_DELLINK after the interface is unregistered from the name table.
if_indextoname fails and handle_nl_link silently returned without calling remove_iface_sockets,
causing old ETCP sockets to accumulate indefinitely.
Fix: fall through to remove_iface_sockets(ifindex) even when if_indextoname fails.
Test: added diagnostics (diag_dump_state, per-category debug levels to file) to verify socket/links cleanup.
- Extend MEMBER_DETAIL_LINK_SIZE to 63B with 20B sock_name per link
- Replace link_count with links_up/links_down (two uint16) in sock snap
- Add BGP hop list collection via topo_node_best_hop_list, post as CHAT_EVT_BGP_INFO
- Add direct RTT summary text via same event
- Rework AccountList detail panel: header + QTextEdit with monospace layout
- Display protocol (TCP/UDP), socket name line, colored dots for connections
- Show [up/down] link counts for self listen sockets
- Remove * prefix from link count display in Android
When etcp_link_close() frees a link while INFLIGHT_PACKET entries
still reference it via last_link, a later ACK causes SIGSEGV in
bbr_main(link->bbr=NULL).
Fix:
- etcp_inflight_nullify_link(): nullifies last_link in both
input_wait_ack and input_send_q for entries pointing to dead_link
- Called from etcp_link_close() before u_free(link)
- Guard etcp_ack_recv: check acked_pkt->last_link->bbr before
accessing BBR/inflight stats
- Also fix bbr leak in etcp_link_close !link->conn path
Verified: 30/30 runs of test_etcp_link_stress pass (was ~50% crash)
Full suite: 71/71 pass
utun_instance_init() did not call instance_init_common(), leaving
instance->client_type and instance->keepalive_interval as 0.
This broke the INIT handshake in V4 protocol: server wrote
keepalive=0 into INIT_RESPONSE, causing negotiate_keepalive
to see 0 and client to receive empty keepalive.
Also added guard: keepalive < 100ms clamped to 100ms with ERROR log.
bb0866b changed close/restart notifications from NULL to loopback_conn().
Remove !conn check — use entry->len==9 alone to distinguish (9 bytes = svc_id+node_id, data always longer).
- Добавлен device_type (1 байт, CLIENT_TYPE_*) в INIT_REQUEST и INIT_RESPONSE
- Добавлен keepalive_interval (2 байта) в INIT_RESPONSE
- STCP handshake расширен с 43 до 46 байт (device_type+keepalive)
- negotiate_keepalive(): оба desktop/server → min, иначе (есть mobile) → max
- keepalive_interval вынесен в UTUN_INSTANCE для использования в handshake
- stcp_client_connect() принимает device_type и keepalive_interval
- peer_device_type/peer_keepalive_interval передаются через stcp_link в ETCP_LINK
10 spammers (1-30ms random) + 2 observers (1-10ms), star topology via hub.
5s of concurrent modifications, then verify all 13 merkle root hashes
are identical and tree is internally consistent (recompute vs stored).
- _recompute_bucket: skip INSERT OR REPLACE when hash unchanged,
also skip DELETE when row doesn't exist (was always writing)
- bg_check interval: 100ms → 5000ms (was writing 10x/sec × 4KB/page
into WAL = ~1.4GB over hours)
- bg_timer: round-robin across channels instead of only first
- bg_timer: periodic WAL checkpoint (every ~60s) to prevent
unbounded WAL growth from any source
- test: sync ms_session struct copy with real definition
- stcp_link: handle race where conn already CLOSED before close
- stcp_client: remove double-free c->free_on_close = cli
- etcp_connections: stcp_link_close on TCP link down
- u_async: round sub-ms timeouts to 1ms to avoid busy-loop
Evgeny
evgeny