When ELBv2Enabled=true:
- Drop lbVMUserData cloud-config; emit KEY=value LBAgentEnv blob instead
- Set SystemInstanceInput.DirectBoot=true; omit ImageID
- Populate NICs []NICConfig (index 0=primary ENI IsDefault, 1=mgmt, 2+=extra ENIs)
- Validate exactly one IsDefault NIC at service layer
When ELBv2Enabled=false: existing PC-machine path unchanged.
Adds buildLBAgentEnv, buildMicrovmNICs, subnetCIDRForIP, subnetGatewayIP helpers.
Tests in arn_test.go cover both flag-on and flag-off paths.
Closes mulga-siv-66
The 5s ec2.start.{lastNode} forward timeout was firing before the target
node finished StartStoppedInstance (cold start = volume rehydrate + QEMU
launch + QMP handshake, ~10-20s on a t3a.nano). The forwarding daemon
then fell back to starting the same instance locally, racing the
still-running launch on the original node. Both attempts derive the same
deterministic tap name from the ENI, so the second one fails with
TUNSETIFF: Device or resource busy.
Surfaced on the pseudo-multinode E2E (run 26006933491) where all three
daemons share one kernel and the collision is unavoidable; real
multi-node hides it because separate kernels can each hold the same tap
name without conflict (but the duplicate Run is still semantically
wrong).
Changes:
* Bump startStoppedForwardTimeout from 5s to 30s so it matches the
awsgw upstream NATS request budget. A slow-but-alive target node
finishes well within this window.
* Distinguish nats.ErrNoResponders (no active subscriber on
ec2.start.{lastNode} — node genuinely down) from a timeout or other
transport error (subscriber existed but didn't respond). Only the
no-responders case justifies a local-fallback Run; everything else
surfaces ServerInternal so the caller retries instead of double-
launching.
* Add an info log on every forward attempt for easier diagnosis.
* feat(predastore): support 2-node distributed mode with RS(1,1)
Lower minimum node threshold from 3 to 2. Emit data=1,parity=1 when
exactly 2 nodes configured — effectively mirroring. 3+ nodes continue
to use RS(2,1).
* fix(ec2): route StartInstances back to original node with node-down fallback
ec2.start used a plain queue group, so NATS could route a restart request to
any daemon — including a capacity-constrained NUC that can't fit a second
instance. This caused InsufficientInstanceCapacity in the Phase 10 stop/start
test on 2-node bare-metal clusters where one node has significantly fewer
resources.
Each daemon now also subscribes to ec2.start.{nodeId} (no queue group). The
queue-group handler peeks at the instance ID, looks up LastNode from the
stopped KV, and forwards to ec2.start.{lastNode}. If the original node does
not respond within 5 seconds (down or partitioned), the handler falls back to
starting locally on whichever node picked up the request — preserving the
existing behaviour for node-failure scenarios.
The direct handler (ec2.start.{nodeId}) always starts locally and never
forwards, preventing routing loops.
* fix(ec2): fall back to local start when original node returns InsufficientInstanceCapacity
The forwarding handler was relaying all responses from the targeted node
unconditionally, including InsufficientInstanceCapacity. This meant that if
the NUC (node2) was the LastNode but had no remaining capacity, the error
propagated to the caller rather than falling back to the local node which may
have room.
Now only success and non-capacity errors are relayed; an InsufficientInstance-
Capacity response from the targeted node triggers a local start attempt, giving
any other node in the cluster a chance to absorb the instance.
* test(e2e): make VPC stop/start restart non-fatal on InsufficientInstanceCapacity
Asymmetric hardware (NUC node2) may not have capacity to restart all
VPC instances. IP-persists-while-stopped is still verified and fails
hard; restart verification is skipped with a warning on capacity errors.
* test(e2e): density-aware capacity guard across bare-metal multi-node suite
Sub-3-node bare-metal clusters with asymmetric/undersized nodes (e.g. a
4-vCPU NUC alongside a 16-vCPU box) can fit only one t3.nano on the
smallest node after the daemon's host reserve. Phases that direct or
indirectly land >1 instance on that node hit InsufficientInstanceCapacity
and previously hard-failed the entire suite.
Add is_density_skip helper and skip_test counter, and wrap every
density-sensitive run-instances / start-instances / wait-for-running call
site (phases 3, 5, 7, 10, 11) so capacity exhaustion on sub-3-node
clusters records a SKIP rather than a FAIL. On 3+ node clusters the
guard short-circuits to false, preserving existing failure semantics.
Summary block now prints Skipped count + names so degraded coverage is
visible in CI logs.
Refs: mulga-siv-72
* test(e2e): parameterise lb SSH key + user via PEER_SSH_KEY / PEER_SSH_USER
Allow run-lb-e2e.sh to run against clusters that don't use tf-user with the
tofu-cluster SSH key (e.g. bare-metal nodes which use spinifex@ + bm-ci key).
Defaults preserve existing behaviour on the tofu-cluster CI path.
Refs: mulga-siv-73
- SSH wait: 60 → 120 attempts; NBD-backed VMs on bare metal take
longer to boot than tofu-cluster VMs with local storage
- Diagnostic console log: add /run/spinifex/console-<id>.log which
is where Spinifex QEMU actually writes serial output
/routez reports outbound routes; with N nodes each node has N-1 routes.
A 2-node cluster has 1 route, not 2. Test was written for 3-node tofu
cluster and failed on 2-node baremetal.
Hardcoded 'for i in 1 2' loops assumed 3+ nodes, accessing NODE_IPS[2]
which doesn't exist on a 2-node cluster. With set -u this is a fatal
unbound variable error. Use seq 1 $((NODE_COUNT-1)) instead.
Phase 8.3 polled VPC instance IPs via two-hop SSH. In single-node macvlan
mode the host has no route to VPC subnets, so SSH always timed out (10 min
of dead retries) even when the VMs were healthy (ALB targets passing health
checks).
Replace with a LaunchTime comparison: the daemon resets LaunchTime to
time.Now() when it relaunches QEMU on recovery, so LaunchTime >= REBOOT_START
is a fast, network-topology-independent reboot confirmation.
GIT_BRANCH was hardcoded to dev, so workflow_dispatch runs ignored the
ref chosen in the UI. Use github.ref_name for dispatch (and any non-cron
trigger), keep dev as the explicit target for scheduled cron runs.
Trixie OVN moved the pid/socket to /var/run/ovn/; older installs
use /var/run/openvswitch/. Probe which exists so the same binary
works on both layouts.
Add SkipFormation bool to firstboot.Config. When set, the cluster
formation block (spx admin init/join) is replaced with a no-op echo,
leaving formation to the provisioning controller.
Wired through: autoinstall reads SPINIFEX_SKIP_FORMATION=1 from env,
install.go passes it to firstboot.Config. spinifex-init.sh parses the
kernel cmdline param and exports it. boot-handler.go hardcodes
SPINIFEX_SKIP_FORMATION=1 so all bare-metal CI nodes skip formation —
bm-bootstrap.sh always runs spx admin init/join with the correct
--nodes N from nodes.json, enabling proper multi-node formation.
Add a step before the E2E run that ensures the Ubuntu 24.04 noble cloud
image is cached at BM_ISO_CACHE. Downloads from ubuntu.com if missing,
prints cached size if present. Passes path via BM_UBUNTU_IMAGE to
bm-bootstrap.sh which SCPs it to the node for --file import.
The --no-nvram flag removed the UEFI NVRAM boot entry, causing firmware to
fall through to PXE on every reboot. Restore original --recheck behaviour
which was confirmed working in manual testing.
grub-probe scans all block devices even without --recheck. On N100 mini-PCs
with eMMC + NVMe, scanning the eMMC causes grub-install to hang. Also skip
NVRAM writes (--no-nvram) since --removable installs to the default EFI path
without needing to register a boot entry in UEFI firmware variables.
--recheck triggers grub to rescan all block devices. On N100 mini-PCs with
eMMC + NVMe, scanning the eMMC causes grub-install to hang indefinitely.
Device maps are generated fresh on each install so --recheck is redundant.
On N100 mini-PCs with eMMC + NVMe, update-grub's os-prober probes all
block devices via the bind-mounted /dev, hanging indefinitely on the eMMC.
GRUB_DISABLE_OS_PROBER=true skips os-prober in the chroot.
Also apply modernize lint fixes to vm test files (atomic.Int32).
* 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
* 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
* installer grub image fixes
* made gpu-passthrough more user facing
* vpcd fixes to set br-wan to promisc for public IP DHCP processes
* fixed vpcd br-wan promisc and veth pair
* ssh keygen fix
* Fix: AMD GPU edge cases
* script for running 8x instances on SMCI computer
* fix: increase disk space on instances
* feat: multi-GPU passthrough support for g7e.12xlarge (2× MI350X)
- vm.VM: GPUPCIAddress string → GPUPCIAddresses []string for multi-GPU
- instancetypes: add g7e.12xlarge (2 GPUs), GPUCountForType(), AMDmi350x model
- instancetypes/generate: use GPUCountForType for GpuInfo.Count and TotalGpuMemoryInMiB
- gpu/manager: Release() loops all pool entries for instanceID (not just first)
- vm/lifecycle: loop over GPUPCIAddresses, emit gpu0/gpu1/... QEMU devices
- daemon/handlers: claim N GPUs in loop for both RunInstances and StartStoppedInstance
- daemon/vm_adapters: ReclaimByAddress loops GPUPCIAddresses; ReleaseGPU checks len > 0
- scripts/launch-gpu-fleet: fixed fleet — 6× g7e.4xlarge + 1× g7e.12xlarge
* feat: per-instance-type disk size in GPU fleet (300 GB single, 600 GB dual)
* fix: multi-gpu instance types
* fix: kill QEMU synchronously when QMP handshake fails to free VFIO device
When startQEMU succeeds but newQMPClientWithHandshake times out, the
instance has not yet been inserted into the manager map so vmMgr.Delete
in the rollback path is a no-op. QEMU keeps running and holds the VFIO
group open. The GPU pool entry gets freed, the next Claim returns the
same PCI address, and every subsequent QEMU fails with 'device or
resource busy'.
Fix: SIGKILL QEMU and wait for it to exit before returning the QMP
error, ensuring the VFIO device is released before the caller recycles
the GPU pool slot.
* fix: quarantine GPU after QEMU launch failure to prevent re-assignment cascade
When QEMU crashes during startup (e.g. pci_irq_handler assertion on a
specific GPU), the pool Release frees that address and the next NATS
request immediately re-claims the same broken device. Add MarkFailed()
to gpu.Manager which sets Available=false for a PCI address after a
launch failure, preventing re-assignment until the daemon restarts.
Call MarkFailed for each GPU address in both RunInstances and
StartStoppedInstance rollback paths after the Release succeeds.
* fixed instance fleet launch script
* remove known hosts and stagger provisioning
* improvements to launch script
* more upgrades
* single instance gpu instance provisioning
* fixed single instance provision
* smci computer demo dashboard
* fixes to launch script
* updated smci dashboard_server
* fixed demo dashboard
* updated dashboard.py
* feat: GPU Passthrough fixes to passthrough process - builds proper AMD or NVIDA ready images, generalised launch script
* feat: gpu-passthrough updating image builder to include gpu-ready ubuntu images
* fixing image builder process
* fixed image builder to remove /tmp/ file before import to free space
* fix: Updated COMMANDS.md to reflect new GPU management admin commands
* Update docs, update internal launch script to work with new SGs, test tweaks
* Updated Docs, fixed internal launch-gpu.sh script, test tweaks
* fxing CI race due to GPU load
joshsiv-mulga