diff options
Diffstat (limited to 'tools/guest-diag/pin_supervisor.sh')
| -rwxr-xr-x | tools/guest-diag/pin_supervisor.sh | 169 |
1 files changed, 169 insertions, 0 deletions
diff --git a/tools/guest-diag/pin_supervisor.sh b/tools/guest-diag/pin_supervisor.sh new file mode 100755 index 0000000..b87d5c1 --- /dev/null +++ b/tools/guest-diag/pin_supervisor.sh @@ -0,0 +1,169 @@ +#!/bin/bash +# SPDX-License-Identifier: BSD-3-Clause +# Copyright (c) 2026 Open-IOV contributors +# pin_supervisor.sh - supervise the host-side pin tools (pin_gr_stream.py and +# semfix_uvmt.py) so a silent death cannot strand every later CUDA client at +# first cudaMalloc. +# +# Both tools write a heartbeat file; this script checks, every 5 s, that each +# process is alive (exact script path) and that its heartbeat mtime is < 75 s. +# A startup grace (120 s, skipped early once the tool log shows it has done +# first useful work) keeps a slow cold start from being restarted. On failure +# it kills the tool (exact path match) and restarts it, appending RESTART +# lines to the supervisor log. After 10 restarts in one hour for a tool it +# logs GIVEUP and exits 1. Only one semfix_uvmt instance is ever started +# (QEMU's QMP socket accepts a single client). +# +# Liveness policy (do not tighten without re-checking the soak evidence): +# * the tools refresh their heartbeats from dedicated loops/threads even +# while idle (pin_gr_stream: 5 s thread; semfix: per loop + around every +# blocking guest-agent/dmesg call), so 75 s never trips on a healthy +# client between phases; +# * a DEAD process is restarted on the next 5 s tick regardless of grace; +# * dead-pid/state recovery (the kill cascade) is handled INSIDE the tools +# (in-band #nopid + idle liveness probe + stall watchdog), NOT by this +# supervisor - a log-based restart here could fire mid-run of a healthy +# client (the m178 lesson). +# +# Usage: pin_supervisor.sh LOGDIR TAG +# Logs: LOGDIR/12-pin-supervisor-TAG.log (supervisor) +# LOGDIR/12-pin_gr_stream-TAG.log, LOGDIR/12-semfix-TAG.log (tools) +set -uo pipefail + +LOGDIR="${1:?usage: pin_supervisor.sh LOGDIR TAG}" +TAG="${2:?usage: pin_supervisor.sh LOGDIR TAG}" + +HERE=$(cd "$(dirname "$0")" && pwd) +PIN="$HERE/pin_gr_stream.py" +SEMFIX="$HERE/semfix_uvmt.py" +PIN_HB="${PIN_GR_HEARTBEAT:-/tmp/opencode/pin_gr_heartbeat}" +SEMFIX_HB="${SEMFIX_HEARTBEAT_FILE:-/tmp/opencode/semfix_heartbeat}" +SUPLOG="$LOGDIR/12-pin-supervisor-$TAG.log" +PIN_LOG="$LOGDIR/12-pin_gr_stream-$TAG.log" +SEMFIX_LOG="$LOGDIR/12-semfix-$TAG.log" +WINDOW=3600 +CAP=10 +STALE=75 +GRACE=120 +mkdir -p "$LOGDIR" /tmp/opencode + +log() { echo "[$(date +%H:%M:%S)] $*" | tee -a "$SUPLOG"; } + +tool_alive() { pgrep -f "python3 $1" >/dev/null 2>&1; } + +hb_age() { # age in seconds of the heartbeat file, 9999 if missing + local f="$1" m + m=$(stat -c %Y "$f" 2>/dev/null) || { echo 9999; return; } + echo $(( $(date +%s) - m )) +} + +kill_tool() { + pkill -f "python3 $1" 2>/dev/null + sleep 1 + if pgrep -f "python3 $1" >/dev/null 2>&1; then + pkill -9 -f "python3 $1" 2>/dev/null + sleep 1 + fi + return 0 +} + +start_pin() { + if tool_alive "$PIN"; then return 0; fi + touch "$PIN_HB" 2>/dev/null || true + ( setsid nohup env SEMFIX_COUNTER_RETRY=10 \ + python3 "$PIN" >> "$PIN_LOG" 2>&1 & ) + pin_started=$(date +%s) + log "START pin_gr_stream" +} + +start_semfix() { + # pgrep exact-count check before start: if an instance already exists + # (ours or a leftover), never start a second one. + if tool_alive "$SEMFIX"; then + log "semfix already running - not starting another" + return 0 + fi + touch "$SEMFIX_HB" 2>/dev/null || true + ( setsid nohup env SEMFIX_SKIP_PB=0 SEMFIX_COUNTER_RETRY=10 \ + python3 "$SEMFIX" 0.5 >> "$SEMFIX_LOG" 2>&1 & ) + semfix_started=$(date +%s) + log "START semfix_uvmt" +} + +prune() { # $1 now, $2 list -> list with entries older than WINDOW dropped + local now="$1" t out="" + for t in $2; do + [ $(( now - t )) -lt "$WINDOW" ] && out="$out $t" + done + printf '%s' "$out" +} + +count_list() { # $1 list -> element count + local n=0 t + for t in $1; do n=$(( n + 1 )); done + echo "$n" +} + +restarts_pin="" +restarts_semfix="" +pin_started=$(date +%s) +semfix_started=$(date +%s) + +restart() { # $1 label, $2 script path, $3 start function + local now list n + now=$(date +%s) + if [ "$1" = pin_gr_stream ]; then + restarts_pin=$(prune "$now" "$restarts_pin") + n=$(count_list "$restarts_pin") + else + restarts_semfix=$(prune "$now" "$restarts_semfix") + n=$(count_list "$restarts_semfix") + fi + if [ "$n" -ge "$CAP" ]; then + log "GIVEUP $1: $n restarts within ${WINDOW}s" + exit 1 + fi + kill_tool "$2" + if [ "$1" = pin_gr_stream ]; then + restarts_pin="$restarts_pin $now" + else + restarts_semfix="$restarts_semfix $now" + fi + log "RESTART $(( n + 1 )) $1 $(date +%Y-%m-%dT%H:%M:%S)" + "$3" +} + +check_tool() { # $1 label, $2 path, $3 heartbeat file, $4 start function, + # $5 tool start TS, $6 tool log file, $7 first-work pattern + local age runtime + age=$(hb_age "$3") + if tool_alive "$2" && [ "$age" -le "$STALE" ]; then + return 0 + fi + if tool_alive "$2"; then + # Startup grace: a cold tool can spend its first ~2 min installing the + # guest scanner or connecting QMP with a lagging heartbeat; only + # enforce staleness once the grace has passed or the log shows the + # tool has already done successful work. + runtime=$(( $(date +%s) - $5 )) + if [ "$runtime" -lt "$GRACE" ] && \ + ! grep -q -- "$7" "$6" 2>/dev/null; then + return 0 + fi + log "$1 heartbeat stale (${age}s) - restarting" + else + log "$1 not running (heartbeat age ${age}s) - restarting" + fi + restart "$1" "$2" "$4" +} + +log "pin_supervisor up (LOGDIR=$LOGDIR TAG=$TAG)" +start_pin +start_semfix +while true; do + sleep 5 + check_tool pin_gr_stream "$PIN" "$PIN_HB" start_pin "$pin_started" \ + "$PIN_LOG" "listening on vsock" + check_tool semfix_uvmt "$SEMFIX" "$SEMFIX_HB" start_semfix \ + "$semfix_started" "$SEMFIX_LOG" "semfix_uvmt: watching" +done |