- Convert tb-recover from timer-triggered oneshot to persistent service
- Browse files
-
bogdan committed 2 weeks ago
tb-recover.timer was triggering the service every 30s, filling logs with spurious "Started" / "Finished" messages even when no recovery was needed. Changed to persistent Type=simple service with continuous loop and 60s monitoring interval. Logs only actionable events (peer failures, recovery attempts), not periodic ticks. - tb-recover.service: Type=oneshot → Type=simple, add Restart+WantedBy - tb-recover.sh: Wrap logic in init_bridge(), handle_missing_interfaces(), monitor_interfaces() functions; add infinite loop with MONITOR_INTERVAL_SEC (default 60s); add SIGTERM trap for graceful shutdown - deploy_tb.sh: Remove timer references; enable tb-recover.service instead - Deleted: tb-recover.timer Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
Showing 4 changed files
with
85 additions
and
84 deletions
+74
-65
projects/thunderbolts/deploy/attempt1/common/sbin/tb-recover.sh
@@ -3,7 +3,6 @@ set -euo pipefail |
||
| 3 | 3 |
|
| 4 | 4 |
BRIDGE="thunderbridge" |
| 5 | 5 |
MTU="65520" |
| 6 |
-FOUND_TB_IFACE=0 |
|
| 7 | 6 |
STATE_DIR="/run/tb-recover" |
| 8 | 7 |
LAST_BOLT_RESTART_FILE="${STATE_DIR}/last_bolt_restart_epoch"
|
| 9 | 8 |
BOLT_RESTART_COOLDOWN_SEC=600 |
@@ -14,9 +13,11 @@ PEER_FAIL_THRESHOLD="${TB_PEER_FAIL_THRESHOLD:-2}"
|
||
| 14 | 13 |
IFACE_CYCLE_COOLDOWN_SEC="${TB_IFACE_CYCLE_COOLDOWN_SEC:-300}"
|
| 15 | 14 |
IFACE_CYCLE_SETTLE_SEC="${TB_IFACE_CYCLE_SETTLE_SEC:-5}"
|
| 16 | 15 |
PING_TIMEOUT_SEC="${TB_PING_TIMEOUT_SEC:-1}"
|
| 16 |
+MONITOR_INTERVAL_SEC="${TB_MONITOR_INTERVAL_SEC:-60}"
|
|
| 17 | 17 |
LOCAL_HOST="$(hostname -s 2>/dev/null || hostname)" |
| 18 | 18 |
|
| 19 | 19 |
mkdir -p "$STATE_DIR" |
| 20 |
+trap "log 'Shutting down tb-recover'; exit 0" SIGTERM SIGINT |
|
| 20 | 21 |
|
| 21 | 22 |
log() {
|
| 22 | 23 |
printf '%s %s\n' "$(date -Is)" "$*" |
@@ -58,6 +59,23 @@ read_counter_file() {
|
||
| 58 | 59 |
peer_ip_for_iface() {
|
| 59 | 60 |
local iface="$1" |
| 60 | 61 |
|
| 62 |
+ # Dynamically resolve peer by looking up the XDomain device name bound to this |
|
| 63 |
+ # interface. The kernel exposes a symlink at /sys/class/net/<iface>/device |
|
| 64 |
+ # pointing to the XDomain service path (e.g. .../1-1.0). Its parent directory |
|
| 65 |
+ # is the XDomain device whose device_name attribute holds the peer hostname. |
|
| 66 |
+ local dev_path xdomain_dev peer_name |
|
| 67 |
+ dev_path="$(readlink -f "/sys/class/net/${iface}/device" 2>/dev/null || true)"
|
|
| 68 |
+ if [ -n "$dev_path" ] && [ -d "$dev_path" ]; then |
|
| 69 |
+ xdomain_dev="$(dirname "$dev_path")" |
|
| 70 |
+ peer_name="$(cat "${xdomain_dev}/device_name" 2>/dev/null || true)"
|
|
| 71 |
+ case "$peer_name" in |
|
| 72 |
+ baobab) printf '%s\n' "192.168.10.91" ; return 0 ;; |
|
| 73 |
+ ebony) printf '%s\n' "192.168.10.92" ; return 0 ;; |
|
| 74 |
+ tapia) printf '%s\n' "192.168.10.93" ; return 0 ;; |
|
| 75 |
+ esac |
|
| 76 |
+ fi |
|
| 77 |
+ |
|
| 78 |
+ # Static fallback (used when sysfs path is not available). |
|
| 61 | 79 |
case "${LOCAL_HOST}:${iface}" in
|
| 62 | 80 |
baobab:thunderbolt0) |
| 63 | 81 |
printf '%s\n' "192.168.10.92" |
@@ -234,83 +252,74 @@ run_nhi_rescan() {
|
||
| 234 | 252 |
return 1 |
| 235 | 253 |
} |
| 236 | 254 |
|
| 237 |
-# Keep the bridge present and up before trying to enslave ports. |
|
| 238 |
-ip link show "$BRIDGE" >/dev/null 2>&1 || ip link add name "$BRIDGE" type bridge || true |
|
| 239 |
-ip link set "$BRIDGE" mtu "$MTU" || true |
|
| 240 |
-ip link set "$BRIDGE" up || true |
|
| 241 |
- |
|
| 242 |
-for path in /sys/class/net/thunderbolt*; do |
|
| 243 |
- [ -e "$path" ] || continue |
|
| 244 |
- IFACE="${path##*/}"
|
|
| 245 |
- FOUND_TB_IFACE=1 |
|
| 246 |
- ip link set "$IFACE" up || true |
|
| 247 |
- ip link set "$IFACE" mtu "$MTU" || true |
|
| 248 |
- ip link set "$IFACE" master "$BRIDGE" || true |
|
| 249 |
- systemctl start "tb-enlist@${IFACE}.service" || true
|
|
| 250 |
-done |
|
| 251 |
- |
|
| 252 |
-# If no thunderbolt netdev exists but a TB domain exists, force a rescan + udev retrigger. |
|
| 253 |
-if [ "$FOUND_TB_IFACE" -eq 0 ] && [ -d /sys/bus/thunderbolt/devices ]; then |
|
| 254 |
- trigger_tb_rescan |
|
| 255 |
- |
|
| 256 |
- # Escalate with cooldown: try PCI NHI remove+rescan to emulate a soft replug. |
|
| 257 |
- sleep 2 |
|
| 258 |
- if ! has_tb_netdev; then |
|
| 259 |
- now="$(date +%s)" |
|
| 260 |
- last="0" |
|
| 261 |
- if [ -f "$LAST_BOLT_RESTART_FILE" ]; then |
|
| 262 |
- last="$(cat "$LAST_BOLT_RESTART_FILE" 2>/dev/null || echo 0)" |
|
| 263 |
- fi |
|
| 255 |
+init_bridge() {
|
|
| 256 |
+ ip link show "$BRIDGE" >/dev/null 2>&1 || ip link add name "$BRIDGE" type bridge || true |
|
| 257 |
+ ip link set "$BRIDGE" mtu "$MTU" || true |
|
| 258 |
+ ip link set "$BRIDGE" up || true |
|
| 259 |
+} |
|
| 264 | 260 |
|
| 265 |
- case "$last" in |
|
| 266 |
- ''|*[!0-9]*) |
|
| 267 |
- last=0 |
|
| 268 |
- ;; |
|
| 269 |
- esac |
|
| 261 |
+handle_missing_interfaces() {
|
|
| 262 |
+ local found_tb_iface=0 |
|
| 263 |
+ |
|
| 264 |
+ for path in /sys/class/net/thunderbolt*; do |
|
| 265 |
+ [ -e "$path" ] || continue |
|
| 266 |
+ found_tb_iface=1 |
|
| 267 |
+ IFACE="${path##*/}"
|
|
| 268 |
+ ip link set "$IFACE" up || true |
|
| 269 |
+ ip link set "$IFACE" mtu "$MTU" || true |
|
| 270 |
+ ip link set "$IFACE" master "$BRIDGE" || true |
|
| 271 |
+ systemctl start "tb-enlist@${IFACE}.service" || true
|
|
| 272 |
+ done |
|
| 270 | 273 |
|
| 271 |
- nhi_last="0" |
|
| 272 |
- if [ -f "$LAST_NHI_RESCAN_FILE" ]; then |
|
| 273 |
- nhi_last="$(cat "$LAST_NHI_RESCAN_FILE" 2>/dev/null || echo 0)" |
|
| 274 |
- fi |
|
| 275 |
- case "$nhi_last" in |
|
| 276 |
- ''|*[!0-9]*) |
|
| 277 |
- nhi_last=0 |
|
| 278 |
- ;; |
|
| 279 |
- esac |
|
| 274 |
+ if [ "$found_tb_iface" -eq 0 ] && [ -d /sys/bus/thunderbolt/devices ]; then |
|
| 275 |
+ trigger_tb_rescan |
|
| 276 |
+ sleep 2 |
|
| 280 | 277 |
|
| 281 |
- if [ $((now - nhi_last)) -ge "$NHI_RESCAN_COOLDOWN_SEC" ]; then |
|
| 282 |
- if run_nhi_rescan "$now"; then |
|
| 283 |
- sleep "$NHI_SETTLE_SEC" |
|
| 284 |
- trigger_tb_rescan |
|
| 285 |
- |
|
| 286 |
- # On newer kernels the first NHI reset can stop at the peer xdomain host |
|
| 287 |
- # node without recreating the matching *.0 network service. |
|
| 288 |
- if ! has_tb_netdev && has_stale_tb_xdomain; then |
|
| 289 |
- retry_now="$(date +%s)" |
|
| 290 |
- if run_nhi_rescan "$retry_now"; then |
|
| 291 |
- sleep "$NHI_SETTLE_SEC" |
|
| 292 |
- trigger_tb_rescan |
|
| 278 |
+ if ! has_tb_netdev; then |
|
| 279 |
+ now="$(date +%s)" |
|
| 280 |
+ last="$(read_epoch_file "$LAST_BOLT_RESTART_FILE")" |
|
| 281 |
+ nhi_last="$(read_epoch_file "$LAST_NHI_RESCAN_FILE")" |
|
| 282 |
+ |
|
| 283 |
+ if [ $((now - nhi_last)) -ge "$NHI_RESCAN_COOLDOWN_SEC" ]; then |
|
| 284 |
+ if run_nhi_rescan "$now"; then |
|
| 285 |
+ sleep "$NHI_SETTLE_SEC" |
|
| 286 |
+ trigger_tb_rescan |
|
| 287 |
+ |
|
| 288 |
+ if ! has_tb_netdev && has_stale_tb_xdomain; then |
|
| 289 |
+ retry_now="$(date +%s)" |
|
| 290 |
+ if run_nhi_rescan "$retry_now"; then |
|
| 291 |
+ sleep "$NHI_SETTLE_SEC" |
|
| 292 |
+ trigger_tb_rescan |
|
| 293 |
+ fi |
|
| 293 | 294 |
fi |
| 294 | 295 |
fi |
| 295 | 296 |
fi |
| 296 |
- fi |
|
| 297 | 297 |
|
| 298 |
- # Secondary fallback with cooldown: restart boltd if interface is still missing |
|
| 299 |
- # and the host actually uses that service. |
|
| 300 |
- if ! has_tb_netdev; then |
|
| 301 |
- if [ $((now - last)) -ge "$BOLT_RESTART_COOLDOWN_SEC" ]; then |
|
| 298 |
+ if ! has_tb_netdev && [ $((now - last)) -ge "$BOLT_RESTART_COOLDOWN_SEC" ]; then |
|
| 302 | 299 |
if systemctl list-unit-files bolt.service >/dev/null 2>&1; then |
| 303 | 300 |
systemctl restart bolt.service || true |
| 304 | 301 |
printf '%s\n' "$now" > "$LAST_BOLT_RESTART_FILE" |
| 305 | 302 |
fi |
| 306 | 303 |
fi |
| 307 |
- fi |
|
| 308 | 304 |
|
| 309 |
- trigger_tb_rescan |
|
| 305 |
+ trigger_tb_rescan |
|
| 306 |
+ fi |
|
| 310 | 307 |
fi |
| 311 |
-fi |
|
| 308 |
+} |
|
| 309 |
+ |
|
| 310 |
+monitor_interfaces() {
|
|
| 311 |
+ for path in /sys/class/net/thunderbolt*; do |
|
| 312 |
+ [ -e "$path" ] || continue |
|
| 313 |
+ assess_peer_health "${path##*/}"
|
|
| 314 |
+ done |
|
| 315 |
+} |
|
| 316 |
+ |
|
| 317 |
+init_bridge |
|
| 318 |
+ |
|
| 319 |
+log "tb-recover monitor started (interval: ${MONITOR_INTERVAL_SEC}s)"
|
|
| 312 | 320 |
|
| 313 |
-for path in /sys/class/net/thunderbolt*; do |
|
| 314 |
- [ -e "$path" ] || continue |
|
| 315 |
- assess_peer_health "${path##*/}"
|
|
| 321 |
+while true; do |
|
| 322 |
+ handle_missing_interfaces |
|
| 323 |
+ monitor_interfaces |
|
| 324 |
+ sleep "$MONITOR_INTERVAL_SEC" |
|
| 316 | 325 |
done |
+6
-1
projects/thunderbolts/deploy/attempt1/common/systemd/system/tb-recover.service
@@ -4,5 +4,10 @@ After=tb-bridge.service bolt.service |
||
| 4 | 4 |
Wants=tb-bridge.service |
| 5 | 5 |
|
| 6 | 6 |
[Service] |
| 7 |
-Type=oneshot |
|
| 7 |
+Type=simple |
|
| 8 |
+Restart=on-failure |
|
| 9 |
+RestartSec=10 |
|
| 8 | 10 |
ExecStart=/usr/local/sbin/tb-recover.sh |
| 11 |
+ |
|
| 12 |
+[Install] |
|
| 13 |
+WantedBy=multi-user.target |
|
+0
-11
projects/thunderbolts/deploy/attempt1/common/systemd/system/tb-recover.timer
@@ -1,11 +0,0 @@ |
||
| 1 |
-[Unit] |
|
| 2 |
-Description=Periodic Thunderbolt recovery probe |
|
| 3 |
- |
|
| 4 |
-[Timer] |
|
| 5 |
-OnBootSec=30s |
|
| 6 |
-OnUnitActiveSec=30s |
|
| 7 |
-AccuracySec=5s |
|
| 8 |
-Unit=tb-recover.service |
|
| 9 |
- |
|
| 10 |
-[Install] |
|
| 11 |
-WantedBy=timers.target |
|
+5
-7
projects/thunderbolts/deploy/attempt1/deploy_tb.sh
@@ -35,7 +35,6 @@ COMMON_UDEV="$BASE_DIR/common/udev/rules.d/90-thunderbolt-net-systemd.rules" |
||
| 35 | 35 |
COMMON_SVC1="$BASE_DIR/common/systemd/system/tb-enlist@.service" |
| 36 | 36 |
COMMON_SVC2="$BASE_DIR/common/systemd/system/tb-bridge.service" |
| 37 | 37 |
COMMON_SVC3="$BASE_DIR/common/systemd/system/tb-recover.service" |
| 38 |
-COMMON_TMR1="$BASE_DIR/common/systemd/system/tb-recover.timer" |
|
| 39 | 38 |
COMMON_BIN1="$BASE_DIR/common/sbin/tb-recover.sh" |
| 40 | 39 |
|
| 41 | 40 |
require() {
|
@@ -80,7 +79,6 @@ deploy_node() {
|
||
| 80 | 79 |
scp -q "$COMMON_SVC1" "${SSH_USER}@${ip}:/etc/systemd/system/tb-enlist@.service"
|
| 81 | 80 |
scp -q "$COMMON_SVC2" "${SSH_USER}@${ip}:/etc/systemd/system/tb-bridge.service"
|
| 82 | 81 |
scp -q "$COMMON_SVC3" "${SSH_USER}@${ip}:/etc/systemd/system/tb-recover.service"
|
| 83 |
- scp -q "$COMMON_TMR1" "${SSH_USER}@${ip}:/etc/systemd/system/tb-recover.timer"
|
|
| 84 | 82 |
scp -q "$COMMON_BIN1" "${SSH_USER}@${ip}:/usr/local/sbin/tb-recover.sh"
|
| 85 | 83 |
|
| 86 | 84 |
echo "==> [$host@$ip] copy NODE config" |
@@ -95,14 +93,14 @@ chmod 0644 /etc/udev/rules.d/90-thunderbolt-net-systemd.rules |
||
| 95 | 93 |
chmod 0644 /etc/systemd/system/tb-enlist@.service |
| 96 | 94 |
chmod 0644 /etc/systemd/system/tb-bridge.service |
| 97 | 95 |
chmod 0644 /etc/systemd/system/tb-recover.service |
| 98 |
-chmod 0644 /etc/systemd/system/tb-recover.timer |
|
| 99 | 96 |
chmod 0755 /usr/local/sbin/tb-recover.sh |
| 100 | 97 |
systemctl daemon-reload |
| 101 | 98 |
udevadm control --reload |
| 102 | 99 |
command -v ifreload >/dev/null 2>&1 && ifreload -a || true |
| 103 | 100 |
systemctl enable --now tb-bridge.service |
| 104 |
-systemctl enable --now tb-recover.timer |
|
| 105 |
-systemctl start tb-recover.service |
|
| 101 |
+systemctl enable --now tb-recover.service |
|
| 102 |
+rm -f /etc/systemd/system/tb-recover.timer |
|
| 103 |
+systemctl daemon-reload |
|
| 106 | 104 |
udevadm trigger --subsystem-match=net --action=add |
| 107 | 105 |
EOF |
| 108 | 106 |
|
@@ -110,7 +108,7 @@ EOF |
||
| 110 | 108 |
ssh $SSH_OPTS "${SSH_USER}@${ip}" bash -s <<'EOF'
|
| 111 | 109 |
set -e |
| 112 | 110 |
systemctl --no-pager --plain --full status tb-bridge.service | sed -n '1,6p' |
| 113 |
-systemctl --no-pager --plain --full status tb-recover.timer | sed -n '1,8p' |
|
| 111 |
+systemctl --no-pager --plain --full status tb-recover.service | sed -n '1,8p' |
|
| 114 | 112 |
systemctl --no-pager --plain --full list-units 'tb-enlist@*.service' | sed -n '1,12p' || true |
| 115 | 113 |
ip -d link show thunderbridge | sed -n '1,3p' |
| 116 | 114 |
bridge link | grep -E 'thunderbolt|thunderbridge' || true |
@@ -120,7 +118,7 @@ EOF |
||
| 120 | 118 |
echo |
| 121 | 119 |
} |
| 122 | 120 |
|
| 123 |
-require "$COMMON_UDEV" "$COMMON_SVC1" "$COMMON_SVC2" "$COMMON_SVC3" "$COMMON_TMR1" "$COMMON_BIN1" |
|
| 121 |
+require "$COMMON_UDEV" "$COMMON_SVC1" "$COMMON_SVC2" "$COMMON_SVC3" "$COMMON_BIN1" |
|
| 124 | 122 |
|
| 125 | 123 |
for h in "${TARGETS[@]}"; do
|
| 126 | 124 |
deploy_node "$h" |