fix(supervision): reap owned session workers after monitor crash

Co-authored-by: GPT-6 <gpt-6@openai.com>
This commit is contained in:
ZeusinaandGPT-6 Bot committed 2026-10-11 09:42:31 +03:00
1 parent d7397ed6cf
commit e9b70b00b6
7 files changed
+206 -10

No files matched your search

+56 -1
View File
@@ -29,6 +29,7 @@ unset ss_password
server_pid=
local_pid=
monitor_pid=
monitor_sid=
zapret_started=0
cooldown_pid=
# Signal only descendants of an owned child, never the container/host namespace.
@@ -40,6 +41,38 @@ pause() {
wait "$cooldown_pid"
cooldown_pid=
}
cleanup_session() {
[ -n "$monitor_sid" ] || return 0
# The verified SID survives leader death and Go's separate worker PGIDs.
# Never sweep all namespace children: Docker exec peers/SS/helpers are not
# in this session. Detached sysv daemons are reconciled/stopped separately.
for session_signal in TERM KILL; do
signal_session "$monitor_sid" "$session_signal" || return 78
session_deadline=$(($(date +%s) + 3))
while :; do
signal_session "$monitor_sid" CHECK || return 78
if [ "$session_live" -eq 0 ]; then
monitor_sid=
return 0
fi
[ "$(date +%s)" -lt "$session_deadline" ] || break
pause 1
done
done
echo 'error: owned monitor session survived cleanup' >&2
return 78
}
capture_monitor_session() {
[ -z "$monitor_sid" ] || return 0
if [ -e "$config_dir/monitor-ready" ]; then
read -r candidate_sid < "$config_dir/monitor-ready" || return 78
case "$candidate_sid" in ''|*[!0-9]*) return 78 ;; esac
[ "$candidate_sid" = "$monitor_pid" ] && [ "$candidate_sid" -gt 1 ] && [ "$candidate_sid" != "$$" ] || return 78
monitor_sid=$candidate_sid
elif proc_identity "$monitor_pid" && [ "$proc_sid" = "$monitor_pid" ] && [ "$proc_sid" -gt 1 ]; then
monitor_sid=$proc_sid
fi
}
# Wait with a watchdog: never let an unresponsive child exceed Docker's grace.
bounded_wait() {
child=$1
@@ -71,13 +104,22 @@ cleanup_cycle() {
cleanup_code=0
# Let the monitor cancel/join scans and restore committed bytes BEFORE stop.
if [ -n "$monitor_pid" ]; then
capture_monitor_session || cleanup_code=78
kill -TERM "$monitor_pid" 2>/dev/null || true
bounded_wait "$monitor_pid" 45
# Cancellation must restore safely even when SS failure requested retry.
case "$waited_code" in 0|143) ;; *) echo "error: monitor cleanup status $waited_code" >&2; cleanup_code=78 ;; esac
[ "$join_timed_out" -eq 0 ] || { echo 'error: monitor join timeout' >&2; cleanup_code=78; }
# SIGTERM may have arrived before the launch acknowledgement was read.
capture_monitor_session || cleanup_code=78
monitor_pid=
fi
# Join/sweep scans even if the monitor was already reaped after SIGKILL.
# A failed sweep vetoes reconciliation and all daemon script sourcing.
if ! cleanup_session; then
cleanup_code=78
zapret_started=0
fi
# Dedicated PID/network namespace: no host/other-target stop.
if [ "$zapret_started" -eq 1 ]; then
# Stop also sources config: clear trust before reconciliation, never
@@ -147,8 +189,21 @@ timeout -k 5 30 /opt/zapret2/init.d/sysv/zapret2 start || exit 78
# fork a monitor we already need to cancel: an early TERM can race interpreter
# startup before its cancellation handler exists. Clean this cycle directly.
{ kill -0 "$server_pid" 2>/dev/null && kill -0 "$local_pid" 2>/dev/null; } || return 75
/usr/local/bin/zapret-autotune managed monitor &
rm -f "$config_dir/monitor-ready"
setsid /bin/sh /opt/managed-target-supervision.sh --monitor "$config_dir" /usr/local/bin/zapret-autotune managed monitor &
monitor_pid=$!
monitor_ack_deadline=$(($(date +%s) + 5))
while [ ! -e "$config_dir/monitor-ready" ]; do
if ! kill -0 "$monitor_pid" 2>/dev/null || [ "$(date +%s)" -ge "$monitor_ack_deadline" ]; then
signal_tree "$monitor_pid" KILL
exit 78
fi
done
capture_monitor_session || exit 78
if proc_identity "$monitor_pid" && [ "$proc_sid" != "$monitor_sid" ]; then
monitor_sid=
exit 78
fi
while kill -0 "$server_pid" && kill -0 "$local_pid" && kill -0 "$monitor_pid"; do
pause 1
done