Files
Audiobuzz 0bfe872828 nvmf: targeted per-version connect (rd.nvmf.nqn) instead of connect-all
Hosts connected with 'nvme connect-all', which attaches EVERY subsystem in the
discovery log. Each host then mounts only its own by-UUID verity namespaces but
leaves a live controller on every other served version's subsystem. Observed via
spdk-release list: more controllers than workstations, each box on every version
(1/3/3/1/3 across 5 versions for ~3 boxes). Beyond wasted connections
(O(hosts x versions)), it defeated the rolling-update drain/reap: no version ever
reached 0 connections while any host was up, so spdk-loader's auto-reap could
never fire and old versions accumulated without bound.

Fix — targeted connect to this build's subsystem only:
- extract-artifacts.sh bakes the per-version SUBNQN (already in verity.json
  .subnqn, matched by the serving side) into the signed UKI cmdline as
  rd.nvmf.nqn=. NQN is version identity, not per-host state, and the cmdline is
  already per-version + immutable-per-signed-UKI, so invariant #1 holds.
- nvmf-connect.sh (test path) uses 'nvme connect -n <nqn>'; legacy UKIs without
  rd.nvmf.nqn fall back to connect-all with a warning so they still boot.
- dot1x-macsec-up.sh (secure path) gains an explicit step-6 targeted connect over
  macsec0 (after it's addressed, per invariant #2), replacing the delegation to
  the dracut nvmf module's discovery connect-all.

Now each host = 1 controller on its own subsystem; versions drain to 0 when their
hosts migrate, so auto-reap works and connections scale O(hosts). Requires a
rebuild + re-sign (cmdline change, invariant #6). BENCH-VERIFY both paths;
confirm the dracut nvmf module isn't also connect-all-ing on the secure path.
CLAUDE.md invariant #2 + TODO #6 updated.
2026-07-03 06:14:08 +00:00

89 lines
4.2 KiB
Bash

#!/bin/bash
# nvmf-connect — ExecStart for nvmf-connect.service (initramfs, TEST MODE).
#
# Why a systemd unit and not a dracut hook: this initramfs is systemd-driven and
# does NOT source dracut initqueue/online hooks when the ip=dhcp link comes up
# (verified on the bench — the online hook installs but never fires: no log, no
# connect). dracut's nvmf module also registers no autoconnect, and a job placed
# on the main initqueue is never re-executed once the loop blocks on events. A
# oneshot unit ordered Before=initrd-root-device.target, with its own
# wait-for-network + connect loop, deterministically connects the NVMe-oF root
# namespace so the by-id device appears and systemd-veritysetup can map
# /dev/mapper/root.
#
# rd.dot1x=0 (open-network test) ONLY — the unit's ConditionKernelCommandLine
# already gates this, but re-check so the script is safe to run directly. The
# secure path connects over macsec0 from dot1x-macsec-up.sh.
type getarg >/dev/null 2>&1 || . /lib/dracut-lib.sh
[ "$(getarg rd.dot1x=)" = "0" ] || exit 0
[ -e /run/dot1x/.nvmf-connected ] && exit 0
_disc=$(getarg rd.nvmf.discover=) # e.g. tcp,192.168.50.10,4420
[ -n "$_disc" ] || { echo "nvmf-connect: rd.nvmf.discover not set" >&2; exit 1; }
_trtype=${_disc%%,*}
_rest=${_disc#*,}
_traddr=${_rest%%,*}
_trsvcid=${_rest##*,}
# Targeted connect: rd.nvmf.nqn pins THIS build's subsystem (baked per-version
# into the signed cmdline by extract-artifacts.sh, matching verity.json .subnqn
# on the target). This replaces a discovery `connect-all`, which attached EVERY
# served version's subsystem to every host — inflating target controllers and
# keeping old versions pinned above zero connections so they never drained/reaped
# (see TODO #6). If rd.nvmf.nqn is absent (a legacy UKI), fall back to connect-all
# so the host still boots — with a warning.
_nqn=$(getarg rd.nvmf.nqn=)
# nvme_tcp is usually auto-loaded; ensure it so /dev/nvme-fabrics exists.
modprobe nvme_tcp 2>/dev/null
# Best-effort boot-splash status for the 'fleet' Plymouth theme (never fatal). By
# the time this runs the ip=dhcp link is up, so mark Network done and NVMe active.
ply() { plymouth update --status="$1" 2>/dev/null || true; }
ply net-done
ply nvme-active
# Keep-alive / reconnect tuning for the ROOT device. These persist across
# switch-root (the in-kernel nvme-tcp controller and its opts survive).
# -k 30 keep-alive timeout (s). The 5s default means a KA every 2.5s; under
# kiosk CPU/memory pressure the nvme-tcp keep-alive work can be delayed
# past 5s, the target sees no KA and disconnects ("keep alive timeout").
# 30s gives a 15s emit interval and far more slack.
# -l -1 ctrl-loss-tmo = retry reconnection forever. This is the rootfs: the
# 600s default makes the host GIVE UP after a blip, after which root I/O
# fails permanently and the box wedges — strictly worse than retrying.
# -c 10 wait 10s between reconnect attempts.
_kato=30
_ctrl_loss_tmo=-1
_reconnect_delay=10
# Networking (ip=dhcp) is brought up in parallel, so poll-and-connect until a
# fabrics namespace appears or we time out. Both connect forms are idempotent
# (re-connecting an already-attached subsystem just errors, harmlessly).
[ -n "$_nqn" ] || echo "nvmf-connect: rd.nvmf.nqn not set — falling back to connect-all (legacy UKI; attaches every served version)" >&2
_i=0
while [ "$_i" -lt 60 ]; do
if [ -n "$_nqn" ]; then
# TARGETED: attach only this build's subsystem (its data + hash namespaces).
nvme connect -t "$_trtype" -a "$_traddr" -s "$_trsvcid" -n "$_nqn" \
-k "$_kato" -l "$_ctrl_loss_tmo" -c "$_reconnect_delay" 2>/dev/null
else
nvme connect-all -t "$_trtype" -a "$_traddr" -s "$_trsvcid" \
-k "$_kato" -l "$_ctrl_loss_tmo" -c "$_reconnect_delay" 2>/dev/null
fi
for _ns in /dev/nvme[0-9]*n[0-9]*; do
if [ -b "$_ns" ]; then
mkdir -p /run/dot1x
: > /run/dot1x/.nvmf-connected
ply nvme-done
echo "nvmf-connect: connected ${_trtype}/${_traddr}/${_trsvcid} ($_ns)"
exit 0
fi
done
sleep 1
_i=$((_i + 1))
done
echo "nvmf-connect: timed out connecting ${_trtype}/${_traddr}/${_trsvcid}" >&2
exit 1