6987b324f1
Some checks failed
With the matrix actually setting session-control, T4 timed a destroyed
master's takeover at:
replace 26s
deny 148s <-- sizing case
disable 21s
`deny` is the case GRACE exists for: the AC refuses the survivor until its own
dead-peer timer frees the dead session. The session poller caught it happening
-- the destroyed router's session stayed in the table while the survivor's
dials appeared and were rejected, twice, before one took at 148s.
148s is well past the lcp-echo-interval(30) x failure(3) = 90s budget that 180
was sized against, leaving 32s of margin. GRACE=300 is ~2x the worst observed.
Treat 148s as a floor, not a worst case: these are idle 2-vCPU VMs, and the AC
shares an OVS bridge with the routers, so `virsh destroy` removes the port and
accel-ppp sees the peer vanish. A real BRAS over DSL never learns our router
died and waits out longer timers of its own.
The cost is stated in the conf: GRACE is also how long an alive-but-unroutable
master holds every VIP before yielding. bond0.53 covers most of that -- a DHCP
lease satisfies the check in seconds -- so GRACE only dominates when PPPoE is
the last path. Kept the health check's fallback in step, since keepalived runs
it with no environment and that number decides mastership if the conf is ever
missing.
check_invariant no longer fails blind on the AC's session count. That count is
only a proxy for "two of our routers dialled", and only while the AC enforces
single-session; under `disable` it does not, so a destroyed router's session
lingers and the count reads 2 with exactly one live router dialled. The real
invariant -- at most one router holds pppoe0 -- is now the failing one, and the
stale session is reported as a WARN rather than silenced, because it still
occupies the slot at a real ISP and is precisely what made `deny` take 148s.
149 lines
7.8 KiB
Bash
Executable File
149 lines
7.8 KiB
Bash
Executable File
#!/bin/sh
|
|
# VRRP health check: may THIS router hold the floating IPs?
|
|
#
|
|
# It may only if it can actually carry the WAN. Without this, VRRP decides
|
|
# mastership purely on whether the peer is still advertising -- so a router with
|
|
# no WAN at all happily takes the VIPs and blackholes the entire LAN's internet
|
|
# while looking perfectly healthy. That is not hypothetical: it is the outage of
|
|
# 2026-09-02, reproduced in labsim.
|
|
#
|
|
# ---------------------------------------------------------------------------
|
|
# The first version of this script asked one question: "do I have an address on
|
|
# a WAN interface". That is correct for a pair where both routers hold WAN all
|
|
# the time. Ours cannot: the 10 gig lease is bound to a cloned MAC and the
|
|
# PPPoE line to a single credential, so the WAN follows mastership (see
|
|
# vrrp-wan-take). Against that design the old check DEADLOCKS --
|
|
#
|
|
# may I be master? -> only if I already have WAN
|
|
# do I have WAN? -> only if I am master
|
|
#
|
|
# -- and the backup sits in FAULT for ever. vyos002 sat exactly there, which
|
|
# meant the pair could not fail over at all: the safety check had quietly
|
|
# removed the redundancy it was protecting.
|
|
#
|
|
# So the question is now asked in the right order: enforce "must have WAN" only
|
|
# on the router that is actually HOLDING the VIPs, and give a new master time to
|
|
# bring the WAN up before judging it.
|
|
# ---------------------------------------------------------------------------
|
|
#
|
|
# exit 0 = eligible for MASTER, non-zero = release and let the peer have it.
|
|
|
|
CONF=/config/vrrp-wan.conf
|
|
[ -r "$CONF" ] && . "$CONF"
|
|
STATE=/run/vrrp-wan
|
|
VIP="${VRRP_WAN_VIP:-192.168.1.1}"
|
|
|
|
# Seconds a new master may go without any WAN. Sourced from vrrp-wan.conf; the
|
|
# fallback is deliberately NOT the old 90. accel-ppp's dead-peer budget is
|
|
# lcp-echo-interval(30) x lcp-echo-failure(3) = 90s, so a hard failover into an
|
|
# access concentrator that does not replace the stale session lands exactly on
|
|
# the boundary: the new master fails its own check, sheds the VIPs, and the peer
|
|
# -- in the same position -- does likewise. Both end in FAULT, which is worse
|
|
# than the outage this check exists to prevent.
|
|
#
|
|
# The fallback tracks vrrp-wan.conf, where the reasoning and the measurements
|
|
# live. Short version: labsim T4 timed a destroyed master's takeover at 26s
|
|
# (replace), 148s (deny) and 21s (disable); `deny` is the sizing case and 180
|
|
# left only 32s over it. Keep the two in step -- keepalived runs this script
|
|
# with no environment, so if vrrp-wan.conf is ever missing THIS number is the
|
|
# one that decides mastership.
|
|
GRACE="${GRACE:-300}"
|
|
|
|
# A deliberate hand-over lever.
|
|
#
|
|
# There is no reliable way to MAKE this pair fail over on demand. VyOS offers
|
|
# only `restart vrrp`, and neither that nor `systemctl restart keepalived` is
|
|
# dependable: with advert_int 1 the peer declares the master dead after ~3.6s,
|
|
# and a restart usually finishes inside that window. Measured in labsim -- the
|
|
# same command moved mastership on one run and not on the next three. A
|
|
# fail-back procedure you cannot trigger on purpose is not a procedure.
|
|
#
|
|
# Failing the health check IS the supported way to shed mastership: the sync
|
|
# group goes FAULT, releases every VIP, and the peer takes over -- the same path
|
|
# a genuine WAN loss takes, so the planned drill exercises the real mechanism
|
|
# rather than a special case.
|
|
#
|
|
# touch /run/vrrp-wan/force-fault hand over within failure-count*interval
|
|
# rm /run/vrrp-wan/force-fault become eligible again (no-preempt keeps
|
|
# it BACKUP until the peer hands back)
|
|
#
|
|
# It lives in /run deliberately: a reboot clears it, so a forgotten drill cannot
|
|
# leave a router permanently ineligible.
|
|
[ -f /run/vrrp-wan/force-fault ] && exit 1
|
|
|
|
# Am I holding the VIPs? Asked of REALITY -- is the management VIP actually on
|
|
# this box -- and not of a /run marker.
|
|
#
|
|
# The marker was the first design and it is unsafe: it is written by the VRRP
|
|
# transition script, and in labsim that script silently failed to run on a
|
|
# promotion (VyOS's keepalived-fifo.py helper stopped delivering while
|
|
# keepalived's own notifies kept working). The router then believed it was
|
|
# backup, passed this check, and sat holding every VIP with no WAN -- the exact
|
|
# outage this script exists to prevent, re-created by trusting the reporter
|
|
# instead of the fact.
|
|
if [ -z "$(ip -4 -o addr show 2>/dev/null | grep " ${VIP}/")" ]; then
|
|
# Clear the grace stamp on the way down, HERE, not only in the reconciler.
|
|
# The reconciler runs every 30s; this runs every 5s. A promotion that
|
|
# inherited a stamp from an earlier mastership scored grace = hours, failed
|
|
# immediately, and took the sync group to FAULT ~5s after passing -- with the
|
|
# peer already faulted, that left BOTH routers in FAULT and the LAN with no
|
|
# gateway. The stamp must belong to the CURRENT mastership or it is worse
|
|
# than useless.
|
|
rm -f "$STATE/since" 2>/dev/null
|
|
exit 0
|
|
fi
|
|
|
|
# Start the grace clock HERE, the moment mastership is first observed.
|
|
#
|
|
# It used to be stamped only by vrrp-wan-reconcile, which runs on a 30s timer --
|
|
# so a freshly promoted master reached this check with no stamp, scored grace=0,
|
|
# failed, and went FAULT before it had any chance to bring the WAN up. The peer
|
|
# then found itself alone with no WAN either and did the same. Observed in
|
|
# labsim: BOTH routers in FAULT, nobody holding the VIPs, the LAN with no
|
|
# gateway at all. That is worse than the outage this script exists to prevent,
|
|
# and it would have hit a REAL failover, not just a drill -- the health check
|
|
# runs every 5s and the reconciler had not yet ticked.
|
|
mkdir -p "$STATE" 2>/dev/null
|
|
[ -f "$STATE/since" ] || date +%s > "$STATE/since"
|
|
|
|
# Master with an address on a WAN interface: healthy.
|
|
#
|
|
# Deliberately NOT "can I reach the internet" and NOT "do I have a default
|
|
# route". During a real ISP outage the route disappears on BOTH routers; a check
|
|
# keyed on that would put both into FAULT, nobody would hold the VIPs, and the
|
|
# LAN would lose inter-VLAN routing too -- turning an internet outage into a
|
|
# total one. A DHCP lease survives an ISP outage, so an address still
|
|
# distinguishes "this box structurally cannot route" from "the internet is down
|
|
# right now", which is the distinction that matters.
|
|
# Any WAN counts. Requiring the 10 gig specifically would fault a healthy master
|
|
# during a genuine 10 gig outage and turn a degraded state into a total one --
|
|
# the same reasoning as the default-route note above. Which one satisfied it is
|
|
# recorded for the operator and the test harness, but does not affect the verdict.
|
|
for ifc in bond0.53 pppoe0; do
|
|
if ip -4 addr show dev "$ifc" 2>/dev/null | grep -q 'inet '; then
|
|
echo "$ifc" > "$STATE/wan" 2>/dev/null
|
|
# Re-stamp on every healthy tick, so the grace window below measures
|
|
# time since this box last DEMONSTRABLY had a WAN.
|
|
date +%s > "$STATE/since" 2>/dev/null
|
|
exit 0
|
|
fi
|
|
done
|
|
rm -f "$STATE/wan" 2>/dev/null
|
|
|
|
# No WAN right now, but there was one within GRACE: ride it out.
|
|
#
|
|
# `since` is re-stamped on every healthy tick, so this measures time since the
|
|
# box last HAD a WAN -- not time since it was promoted. Measuring from promotion
|
|
# was wrong in a way that only shows up on an established master: after hours of
|
|
# uptime `now - since` far exceeds any grace, so the first moment bond0.53 went
|
|
# down and pppoe0 was mid-redial, the master failed its own check, shed every
|
|
# VIP, and the peer -- inheriting the same WAN outage -- did the same. Observed
|
|
# in labsim: taking the 10 gig down flapped the pair instead of falling back to
|
|
# PPPoE. A WAN gap must be survivable wherever it happens, not only just after a
|
|
# promotion.
|
|
since=$(cat "$STATE/since" 2>/dev/null || echo 0)
|
|
[ $(( $(date +%s) - since )) -lt "$GRACE" ] && exit 0
|
|
|
|
# Master, past grace, still no WAN: release. This is the 2026-09-02 case.
|
|
exit 1
|