NBDKit exiting within 1s was only treated as an error if exit code != 0.
A clean exit (code=0) fell through to Mounted=true, leaving QEMU with a
dead NBD socket and VMs that never boot post-reboot.
Treat any premature exit as failure regardless of exit code.
manager_deps_test.go: drop TestDeps_VolumeMounterSatisfiesInterface,
TestDeps_StateStoreSatisfiesInterface, TestDeps_StateStoreErrorPropagates,
TestSetDeps_OverridesDeps, TestManagerHooks_InitiallyNil — they assert a
fake recorded its own input, that a fake returns an error literally
constructed in the test, or Go zero-value/direct-assignment semantics.
Compile-time interface checks already prove satisfaction.
state_test.go: replace TestEC2StateCodes_AllStatesHaveMapping and
TestEC2StateCodes_CorrectValues — both compared the production map to
its own literal. Now assert against an independent AWS contract using
ec2.InstanceStateName* SDK enum constants paired with the
AWS-documented numeric codes; a typo in production no longer
replicates into the test.
watchdog_test.go: replace TestStartPendingWatchdog_CtxCancelStopsGoroutine's
time.Sleep(10ms) "assertion" with goleak.VerifyNone. The previous form
passed even if the goroutine never returned on cancel — the harness
reaped it on test exit.
* feat: add GPU discovery package (mulga-992)
Adds spinifex/gpu package with sysfs-based GPU enumeration, IOMMU group
member lookup, and a PCI device ID model table. Vendor-specific enrichment
via nvidia-smi is attempted opportunistically and falls back to the built-in
table when unavailable. All discovery and IOMMU logic is covered by unit tests
using mock sysfs trees.
* feat: add VFIO bind/unbind and GPU pool manager (mulga-993)
Adds gpu.BindVFIO/UnbindVFIO using sysfs driver_override for per-device
driver switching, and gpu.Manager which owns the claim/release lifecycle
for the GPU pool including full IOMMU group binding, rollback on partial
failure, and marking entries unavailable when rebind fails after VM
termination.
* feat: add G5 instance type generator with GPU model support (mulga-994)
- Add GPUModel type, NVIDIAa10g constant, and GPUModelForVendorDevice lookup
- Add generateGPUTypes to emit g5.* types with GpuInfo populated
- Add IsGPUType helper for consumers filtering GPU-capable types
- Update DetectAndGenerate to accept gpuModels []GPUModel parameter
* feat: integrate GPU passthrough into daemon lifecycle (mulga-995)
- Add gpu_passthrough config flag; discover GPUs and build manager at startup
- Gate g5 instance types on GPU pool availability via canAllocateCount
- Claim GPU (VFIO bind) in LaunchInstance; inject vfio-pci QEMU device
- Release GPU (VFIO unbind, rebind to host driver) in stopInstance
- Re-claim GPU on daemon restart when QEMU process is still alive
- Add UseUEFI / OVMF support for x86_64 VMs (optional, falls back to SeaBIOS)
- Fix Claim to use stored OriginalDriver when device is already vfio-pci bound
* feat: add GPU fields to NodeStatus and spx top nodes (mulga-996)
- Add TotalGPUs/AllocGPUs to NodeStatusResponse (zero on non-GPU nodes)
- Populate from gpuManager in handleNodeStatus
- Show GPU (used/total) column in spx top nodes; "-" when no GPUs present
* feat: Ubuntu 24.04 GPU AMI build pipeline (mulga-997)
Extends build-system-image.sh to support Ubuntu alongside Alpine.
Key differences handled: partition table (parted resizepart p1),
bind mounts for DKMS (/proc /sys /dev), apt vs apk, systemctl vs
rc-update. DISTRO defaults to "alpine" for backward compatibility.
Adds ubuntu-gpu.conf manifest and ubuntu-gpu-setup.sh chroot script
that installs nvidia-headless-550-server with DKMS pre-built against
a pinned linux-image-generic kernel, blacklists nouveau, and prevents
unattended-upgrades from invalidating the pre-built module.
* Fixed GPU passthrough and added test script
* feat: always-probe GPU at startup, surface capability in NodeStatusResponse
- Add gpuProbeResult struct and probeGPU() which runs at daemon startup
regardless of gpu_passthrough config. Checks sysfs for GPUs, IOMMU
groups, and vfio_pci module presence.
- Store gpuProbe on Daemon struct; split NewDaemon into probe (always)
and activate (opt-in) phases.
- Populate GPUCapable, GPUPassthrough, GPUModels on NodeStatusResponse
so operators can see hardware state before enabling passthrough.
- Update 'spx top' GPU column: '-' = no hardware, '0/N*' = capable but
passthrough not active, 'used/total' = passthrough active.
* feat: SIGHUP hot-reload for GPU passthrough config (mulga-1001)
- Export GenerateGPUTypes for use by the reload path
- Add ResourceManager.reloadGPUTypes: mutates instanceTypes map in-place
(preserves references held by instanceService) and updates NATS
subscriptions without full daemon restart
- Separate SIGHUP from SIGTERM/SIGINT: setupReload() listens for SIGHUP
in a context-aware loop; setupShutdown no longer intercepts SIGHUP
- Add reloadConfig() + applyGPUConfig(): re-reads spinifex.toml on signal,
handles false→true (re-probe + activate) and true→false (refuse if
instances running, otherwise deactivate) transitions
* feat: spx admin gpu status/enable/disable commands (mulga-1000)
- admin.SetGPUPassthrough: Go implementation of the idempotent TOML write
(replaces the inline Python heredoc in spx-test-gpu.sh)
- spx admin gpu status: shows GPU hardware, IOMMU/vfio state, passthrough
activation, and advertised g5 instance types via NATS NodeStatusResponse
- spx admin gpu enable: checks prerequisites, writes TOML, sends SIGHUP,
polls until daemon confirms GPUPassthrough=true
- spx admin gpu disable: refuses while GPU instances are running, otherwise
writes TOML, sends SIGHUP, polls until confirmed
- spx-test-gpu.sh: replace Python TOML heredoc with spx admin gpu enable
* fixed admin
* fix: filter Intel iGPU from passthrough pool, use lspci for consumer GPU names
Discover() now skips non-NVIDIA/AMD vendors so Intel iGPUs never enter
the passthrough pool. enrichFromLspci() uses the system PCI ID database
via lspci -vmm to populate the model name for any card not in the
hardcoded table, without needing nvidia-smi or driver tools.
* fix: grant CAP_SYS_ADMIN and vfio device access to spinifex-daemon
GPU passthrough requires writing to /sys/bus/pci/drivers/vfio-pci/unbind
on instance stop. The kernel gates this on CAP_SYS_ADMIN regardless of
file permissions. QEMU subprocesses also need char-vfio rw access to
open /dev/vfio/N for the passed-through device.
* fixing gpu bining permissions
* working state of gpu passthrough on phyiscal box
* installer grub image fixes
* made gpu-passthrough more user facing
* fixed banner
* fixed banner
* banner fix
* gpu passthrough docs
* preflight tweak
* vpcd fixes to set br-wan to promisc for public IP DHCP processes
* fixed vpcd br-wan promisc and veth pair
* ssh keygen fix
* fixes after running /review-branch
* feat: re-wire GPU claim and device injection after vm.Manager refactor
GPU passthrough claim (VFIO bind) and QEMU device injection were lost
when dev's vm.Manager refactor (#185/#186) moved StartInstance/
CreateQMPClient into vm/lifecycle.go. Re-add both in their correct
new homes:
- daemon_handlers_instance.go: claim GPU before vmMgr.Run() in both
RunInstances and StartStoppedInstance handlers; release on failure
- vm/lifecycle.go startQEMU(): inject vfio-pci device with x-vga flag
derived from instance.GPUXVGAEnabled when GPUPCIAddress is set
* fix: Create spinifex user ~/.ssh and keys in firstboot
---------
Co-authored-by: Ben Duncan <5178243+benduncan@users.noreply.github.com>
Co-authored-by: Julian Sommer <115516918+juliansommer@users.noreply.github.com>
Scheduled runs were using github.ref_name, which equals the default branch
(main) — so the nightly was testing main. Hardcode GIT_BRANCH=dev so the
build-install-artifacts.sh + tofu provisioning target dev's HEAD instead.
GitHub always loads scheduled-workflow files from the default branch, so
this change must stay in sync on main for the schedule to use it.
The post-reboot guest uptime check was a single-shot ssh with a 10s connect
timeout. After the host comes back, guest QEMU + cloud-init + sshd inside
each guest still need a few seconds to settle, so the check raced them.
Poll up to GUEST_SSH_SECS (default 180s) per guest, mirroring the host
TCP/22 poll. Run 25329248329 cell-18 reproduced this — ALB targets went
healthy ~2s after the second SSH attempt failed.
When a function returning non-zero is captured via command substitution
under `set -euo pipefail`, the assignment exits the script before the
following `if [ -z "$var" ]` diagnostic can run. CI failures showed up
as a sudden jump to the EXIT trap with no error message.
- run-multinode-e2e.sh: wrap get_remote_ssh_port in `if !`, bump retry
10 -> 30, add dump_guest_ssh_diagnostics on miss
- run-pseudo-multinode-e2e.sh, run-e2e.sh: wrap get_ssh_port in `if !`
- run-lb-e2e.sh: add `|| VAR=""` fallback on pick_peer
Add WAN bridge prerequisite callout to single-node and multi-node
install guides. Bootstrap installer assumes br-wan exists; running
on a host whose default route is on a bare NIC leaves install in a
non-working state. Auto-provisioning will land in next release.
Both services used one-shot utils.ConnectNATS in launchService. On ISO
firstboot, when nats.service is starting in parallel, the dial can race
ahead of NATS being ready and exit non-zero. systemd's
Restart=on-failure with default StartLimitBurst=5 in 10s burns through
the budget in ~25s and the unit enters failed permanently — DHCP
manager never starts, vpc.dhcp.acquire requests hit no-responders and
the install hangs nondeterministically.
Match daemon/awsgw, which already use ConnectNATSWithRetry (5min
exponential backoff). The newDHCPManagerWithRetry guard (mulga-siv-40)
only helps after the dial succeeds.
Conflict resolution:
- external_ipam_dhcp.go: take dev deletion (consolidated into dhcp package)
- external_ipam.go: take dev's dhcp.RequestAcquire call
- external_ipam_test.go: drop duplicate ObtainDHCPLease/ReleaseDHCPLease tests
- daemon.go: drop dead gwMAC block (GatewayMAC field not in dev's ExternalPoolConfig)
Playbooks for preflight / teardown / install / reset covering the same
ground as dev-install.sh, dev-setup.sh, reset-dev-env.sh. Existing bash
scripts untouched — decide which to keep after parity testing.
- roles/teardown + teardown catalog README track every known state
source; add a task + row when a leak is reported.
- roles/init gates smoketest on awsgw readiness (describe-regions poll).
- roles/smoketest is dup-tolerant, retry-wrapped, auto-terminates the
launched instance.
- dev-install.yml skips build/install/init when a healthy environment
is already present (4-signal health check), always runs smoketest.
- dev-reset.yml captures region/az/external-mode scalars from the
existing spinifex.toml and replays them into init.
Plan: docs/development/improvements/ansible-dev-lifecycle.md (mulga).
Bead: mulga-siv-9.
* fix(installer): emit kernel cmdline to ttyS0 alongside tty0 (mulga-siv-33)
GRUB_CMDLINE_LINUX in /etc/default/grub only listed console=tty0, so the
installed kernel printed nothing on serial after the GRUB menu. Headless
installs (CI, racks with serial-only IPMI, remote-dev SSH-to-QEMU) saw a
black screen forever and had no path to debug boot failures.
Append console=ttyS0,115200 — the last `console=` on the cmdline becomes
the system console, so ttyS0 wins on serial-only hardware while tty0
still receives output on local displays.
Plan: docs/development/bugs/iso-public-ip-boot-race.md
* fix(ipam): widen vpc.dhcp.acquire retry budget to 60 s (mulga-siv-34)
ObtainDHCPLease retried only 5 attempts × 1 s before returning a hard
error. On a cold ISO boot, NATS comes up well before vpcd's reconcile
leader-election + dhcp.Manager subscribe (typical 5–30 s, pathological
~60 s during full reconcile), so daemon's NewExternalIPAM races vpcd's
dhcp.acquire subscription and fails permanently — d.externalIPAM stays
nil for the daemon's lifetime, daemon_handlers_instance.go silently
skips all auto-public-IP allocation, and instances launch with no
public IP, no dnat_and_snat, and no NAT event ever published.
Bump maxAttempts to 60 with 1 s spacing; expose both as package vars
so tests can shrink the budget for fast timeout-path coverage. Pure
widening of the existing retry loop — no behaviour change on the
success path, only the worst-case wallclock when vpcd never comes up.
The dominant cause of slow vpcd starts (mulga-siv-33: systemd-networkd
not enabled in the ISO → veth pair vanishes on reboot → vpcd
crashloops) is fixed alongside this. A 60 s budget fits comfortably
above all observed healthy-startup times.
Plan: docs/development/bugs/iso-public-ip-boot-race.md
* fix(setup-ovn): declare Linux WAN bridge as NetDev so networkd-managed veth resolves Bridge= on reboot (mulga-siv-33)
The veth-mode reboot persistence drop-ins added in mulga-998.b were
incomplete: setup-ovn.sh wrote `15-spinifex-veth-wan.network` with
`Bridge=br-wan`, but never declared br-wan itself to networkd.
br-wan is created by `networking.service` (ifupdown via
`/etc/network/interfaces`, generated by `install.go writeNetworkConfig`),
not networkd. networkd's `Bridge=` directive requires the target to be a
known NetDev — when it isn't, networkd silently logs
`br-wan NetDev could not be found, ignoring assignment` and the veth
ends up with no master. vpcd's bridge-mode sanity check then fails on
every reboot.
Fix: also write `14-spinifex-br-wan.netdev` declaring br-wan as
`Kind=bridge`. networkd matches the existing kernel bridge by name+kind
(the `Failed to create netdev: File exists` log is harmless), and
`Bridge=br-wan` resolves on next boot. ifupdown still owns DHCP and the
IP — networkd just needs the NetDev declaration to enslave the veth.
The write is gated to avoid clobbering operator-managed networkd config:
if any other `.netdev` under `/etc/systemd/network`,
`/run/systemd/network`, or `/usr/lib/systemd/network` already declares
the WAN bridge by name, the write is skipped with a log line pointing
at the operator file. Cleanup hook also removes the new file when
switching away from veth mode.
Verified end-to-end via QEMU repro: clean install + first boot + reboot,
veth-wan-br master persists across reboots, vpcd starts cleanly.
* fix(vpcd/dhcp): enable AF_PACKET promisc on bridge for DORA (mulga-siv-33)
Real-world DHCP servers commonly ignore the BOOTP BROADCAST flag and
reply with a unicast OFFER addressed to the synthetic per-tenant chaddr
(e.g. 02:f3:...). A Linux bridge floods unknown unicast to slave ports
but does NOT pass it up to the bridge netdev unless IFF_PROMISC is set,
so vpcd's AF_PACKET socket on br-wan misses the OFFER and DORA times
out at 15s ("unable to receive an offer: context deadline exceeded").
Dev/CI didn't hit this:
- CI uses spx_wan_mode=static — no DORA via vpcd.
- QEMU dev DHCP servers (SLIRP/dnsmasq) honor the BROADCAST flag, so
their replies are L2 broadcast which the bridge always passes up.
Replace the three nclient4.New(iface, ...) call sites in Acquire/Renew/
Release with nclient4.NewWithConn(...) wrapping a custom AF_PACKET
DGRAM conn that calls SetPromiscuous(true). Promisc is per-socket
(per-DORA, ~15s) and refcounted by the kernel — auto-clears on close.
* revert(vpcd/dhcp): drop AF_PACKET promisc wrapper, restore nclient4.New
ecfbd4f swapped nclient4.New for openPromiscBridgeConn + NewWithConn to
catch unicast OFFER replies that bypass the bridge netdev. On dev hosts
where upstream honors the BOOTP BROADCAST flag, the wrapper regressed
DORA — public-IP allocation failed where main works. Reverting until
the promisc path is validated against both broadcast- and unicast-OFFER
servers.
* fix(setup-ovn): mark br-int/br-ext Unmanaged for systemd-networkd (mulga-siv-37)
Fresh ISO install on Trixie+systemd-networkd left the OVS internal netdevs
admin-down: networkd has no matching .network so it tags them as unmanaged
but never brings them up. OVS dataplane works but ovn-controller flow
programming and any tooling that probes link state misbehaves.
Drop /etc/systemd/network/05-spinifex-ovs-internal.network with
Unmanaged=yes so OVS itself owns the link state — `ovs-vsctl add-br`
already runs the netdev up at create time.
* fix(vpcd): allocate gateway LRP IP from pool in centralized NAT (mulga-siv-36)
In veth/macvlan modes the gateway LRP carries all egress, so it must hold
a WAN-subnet IP — otherwise the upstream router silently drops ARP for
the WAN nexthop (RFC 826) and the default route never resolves. The
hardcoded 169.254.0.1/30 was correct for distributed NAT but regressed
centralized NAT (introduced under mulga-siv-26).
Add gw_lrp_range_start/end to ExternalPool. In centralized mode at IGW
attach the gateway LRP gets a fresh IP from that range, persisted as
external_ids:spinifex:gateway_ip on the LRP for idempotent reconcile and
retrofit. Direct mode keeps link-local. Per-VM EIP allocator skips the
gw range; ValidatePoolConfig rejects overlap with the EIP range.
Adds TestTopologyHandler_IGWAttach_CentralizedNATAllocatesGwLrpIP and
extends RetrofitAllGatewayPortNetworks to migrate stale link-local LRPs
to allocated pool IPs at startup.
* fix(vpcd): drop link-local fallback, auto-derive gw_lrp_range (mulga-siv-36)
In centralized NAT (veth/macvlan) the gateway LRP must hold a WAN-subnet
IP — link-local cannot ARP-resolve the upstream nexthop, so the upstream
router silently drops egress. Hardware deployment confirmed the previous
fallback to 169.254.0.1/30 was always reached when gw_lrp_range was not
explicitly configured.
Auto-derive a 16-IP range at the top of the WAN subnet (broadcast - 16
.. broadcast - 1) when gw_lrp_range is unset; shift below RangeStart on
overlap with the per-VM EIP range; clamp inside the subnet; guard the
upstream gateway IP. IGW attach now errors out instead of falling back
to link-local when allocation fails — surfaces misconfiguration loudly.
Adds focused unit tests for gwLrpRange, ValidatePoolConfig and
nextAvailableExternalIP gw-lrp skip; updates the multi-VPC and
PoolWithGatewayIP regression tests.
* feat(vpcd): DHCP-acquire gateway LRP IPs (mulga-siv-38)
Per-VPC OVN gateway LRPs now obtain their WAN-subnet IP via vpc.dhcp.acquire
when the pool is source="dhcp", instead of carving from a static gw_lrp_range.
The upstream DHCP server becomes the single arbiter of LAN address space; a
cluster grows by N nodes without operator-side range carving in spinifex.toml.
- Consolidate ObtainDHCPLease/ReleaseDHCPLease into the dhcp package as
RequestAcquire/RequestRelease so vpcd and the daemon-side ExternalIPAM
share one helper. handlers/ec2/vpc keeps its existing ENI/EIP DHCP path.
- TopologyHandler dispatches by pool source: DHCP pools call DHCPManager
via NATS (client-id "gw-lrp-<vpcId>", idempotent on retry); static pools
keep the siv-36 gw_lrp_range / auto-derive allocator. Direct mode and
no-pool deployments still use link-local.
- IGW detach issues vpc.dhcp.release so the upstream server reclaims the
IP. Best-effort — server-side expiry is the safety net.
- Tests: TestTopologyHandler_IGWAttach_DHCPGwLrpIP (3 VPCs share one DHCP
pool, distinct IPs, idempotent attach, release on detach),
IGWAttach_DHCPAcquireFailureAborts (no link-local fallback when DHCP
errors), TestPrefixFromMask. Existing static / auto-derive tests
preserved.
* fix(vpcd): subscribe DHCPManager before reconcile to fix boot race (mulga-siv-38)
DHCPManager.Subscribe ran AFTER ReconcileFromKV / RetrofitAllGatewayPortNetworks,
so retrofit's vpc.dhcp.acquire requests for existing IGWs hit "no responders" and
burned the 60-attempt × 1 s cold-boot retry budget per VPC before failing. On a
dev box with one bootstrap VPC the symptom is a 60 s vpcd startup hang followed
by "allocate gw LRP IP failed", and the bootstrap IGW topology missing from OVN
until the next reconcile pass.
Move dhcpManager.Subscribe ahead of the two reconcile passes so the subscription
is live before any topology code can fan out an acquire.
* fix(vpcd): also subscribe DHCPManager before bootstrap reconcile pass
Previous fix moved DHCPManager.Subscribe ahead of ReconcileFromKV but
missed Pass 1 (the bootstrap-config Reconcile that runs even earlier).
On a dev box with a source="dhcp" pool and a bootstrap IGW, that earlier
pass still hit "no responders" and burned the 60-retry cold-boot budget
before the Manager came up. Move both reconcile passes after Subscribe.
* fix(vpcd): RFC 2131 §4.1 DORA exponential backoff (mulga-siv-39)
DHCPManager.Acquire previously did one DISCOVER and gave up after 15s
because the outer ctx.WithTimeout matched the per-attempt timeout —
nclient4's WithRetry(3) could never fire. Real DHCP servers (consumer
routers, ISC dhcpd, dnsmasq) silently drop OFFERs under load; the spec
mandates retransmission with exponential backoff, and a single lost
DISCOVER was killing legitimate boots.
Move retransmission into DHCPManager.acquireWithBackoff with a
4s/8s/16s/32s schedule (±1s jitter, ~60s worst case). NClient4Client
becomes single-attempt; manager owns retry policy. Default
acquireTimeout lifts from 15s to 90s (sum of schedule + slack).
* fix(vpcd): retry NewDHCPManager on JetStream-not-ready (mulga-siv-40)
NewDHCPManager creates the spinifex-dhcp-leases KV bucket via
utils.GetOrCreateKVBucket → js.CreateKeyValue with no retry. After the
mulga-siv-38 launchService reorder, this runs before any other JetStream
user, so on multi-node cold-boot vpcd hits "JetStream system temporarily
unavailable" and exits before cluster quorum forms. Pseudo-multinode CI
calls start_node_services sequentially, so node2's vpcd failure aborts
the script before node3 ever launches and quorum then never forms.
Wrap NewDHCPManager in newDHCPManagerWithRetry (500ms→10s, 5 min cap,
ctx-cancellable) mirroring daemon.initJetStream so vpcd waits as
patiently as the daemon for the cluster to come up.
* fix(daemon): idempotent DetachVolume QMP path (mulga-siv-41)
DetachVolume's QMP sequence was not retry-safe. First call: device_del
succeeds, blockdev-del fails with "Node nbd-vol-X is in use" because NBD
client teardown / in-flight I/O still hold the block node. Handler bails
ServerInternal "leaving volume state intact" but the guest device is
actually gone. AWS-CLI retries: device_del returns DeviceNotFound,
handler bails again. After 2 retries aws cli exits 254 — Single-Node
Phase 8 Step 3 reproducibly fails this way.
Phase 1: treat DeviceNotFound on device_del as already-detached, log and
continue. Detach is now idempotent across retries.
Phase 2: tryBlockdevDel polls blockdev-del up to 20×detachDelay on "in
use" errors so the I/O drain can complete instead of failing the whole
detach. Other QMP errors fall through immediately as before.
DetachVolume's QMP sequence was not retry-safe. First call: device_del
succeeds, blockdev-del fails with "Node nbd-vol-X is in use" because NBD
client teardown / in-flight I/O still hold the block node. Handler bails
ServerInternal "leaving volume state intact" but the guest device is
actually gone. AWS-CLI retries: device_del returns DeviceNotFound,
handler bails again. After 2 retries aws cli exits 254 — Single-Node
Phase 8 Step 3 reproducibly fails this way.
Phase 1: treat DeviceNotFound on device_del as already-detached, log and
continue. Detach is now idempotent across retries.
Phase 2: tryBlockdevDel polls blockdev-del up to 20×detachDelay on "in
use" errors so the I/O drain can complete instead of failing the whole
detach. Other QMP errors fall through immediately as before.
NewDHCPManager creates the spinifex-dhcp-leases KV bucket via
utils.GetOrCreateKVBucket → js.CreateKeyValue with no retry. After the
mulga-siv-38 launchService reorder, this runs before any other JetStream
user, so on multi-node cold-boot vpcd hits "JetStream system temporarily
unavailable" and exits before cluster quorum forms. Pseudo-multinode CI
calls start_node_services sequentially, so node2's vpcd failure aborts
the script before node3 ever launches and quorum then never forms.
Wrap NewDHCPManager in newDHCPManagerWithRetry (500ms→10s, 5 min cap,
ctx-cancellable) mirroring daemon.initJetStream so vpcd waits as
patiently as the daemon for the cluster to come up.
DHCPManager.Acquire previously did one DISCOVER and gave up after 15s
because the outer ctx.WithTimeout matched the per-attempt timeout —
nclient4's WithRetry(3) could never fire. Real DHCP servers (consumer
routers, ISC dhcpd, dnsmasq) silently drop OFFERs under load; the spec
mandates retransmission with exponential backoff, and a single lost
DISCOVER was killing legitimate boots.
Move retransmission into DHCPManager.acquireWithBackoff with a
4s/8s/16s/32s schedule (±1s jitter, ~60s worst case). NClient4Client
becomes single-attempt; manager owns retry policy. Default
acquireTimeout lifts from 15s to 90s (sum of schedule + slack).
Previous fix moved DHCPManager.Subscribe ahead of ReconcileFromKV but
missed Pass 1 (the bootstrap-config Reconcile that runs even earlier).
On a dev box with a source="dhcp" pool and a bootstrap IGW, that earlier
pass still hit "no responders" and burned the 60-retry cold-boot budget
before the Manager came up. Move both reconcile passes after Subscribe.
Julian Sommer