KVM: Hot-plugged NIC gets non-sequential PCI slot causing unpredictable interface naming in guest
- Dominant language
- Java
- Stars
- 3.1k
- Forks
- 1.4k
- Avg merge
- 6d 19h
- Merged PRs (30d)
- 32
Description
## Description
When hot-plugging a NIC to a running KVM VM via CloudStack, the new NIC receives a non-sequential PCI slot address, resulting in unpredictable interface naming in the guest OS (e.g., `ens9` instead of the expected `ens5`). This causes guest network configuration to fail because CloudStack and cloud-init expect sequential interface naming.
## Root Cause
In `LibvirtPlugNicCommandWrapper.java`, the `execute()` method calls `vifDriver.plug()` which creates an `InterfaceDef` without setting a PCI slot (line 64). When `vm.attachDevice()` is called, libvirt auto-assigns the next free PCI slot. Since non-NIC PCI devices (virtio-serial controller, virtio-disk, memballoon, watchdog) occupy slots immediately after the existing NICs, the hot-plugged NIC gets a much higher slot number.
### Typical PCI slot layout for a CloudStack KVM VM:
| Slot | Device |
|------|--------|
| 0x01 | USB + IDE controllers |
| 0x02 | Video (cirrus) |
| 0x03 | NIC net0 → guest sees `ens3` |
| 0x04 | NIC net1 → guest sees `ens4` |
| 0x05 | virtio-serial controller |
| 0x06 | virtio-disk |
| 0x07 | memballoon |
| 0x08 | watchdog |
| 0x09 | **Hot-plugged NIC → guest sees `ens9` (should be `ens5`)** |
Note: `LibvirtReplugNicCommandWrapper` already handles this correctly by preserving the PCI slot when re-plugging (`interfaceDef.setSlot(oldPluggedNic.getSlot())`), but `LibvirtPlugNicCommandWrapper` has no slot assignment logic.
## Evidence (from production CloudStack 4.20)
**Before fix** — hot-plugged NIC at PCI slot 0x09:
```
$ virsh dumpxml i-2-1681-VM | grep -A6 'net2'
$ virsh qemu-agent-command i-2-1681-VM '{"execute":"guest-network-get-interfaces"}'
"name": "ens3", "hardware-address": "02:01:01:76:01:35"
"name": "ens4", "hardware-address": "02:01:03:15:00:01"
"name": "ens9", "hardware-address": "02:01:03:4c:00:08" ← non-sequential
$ ip a (inside guest):
4: ens9: mtu 1500 qdisc noop state DOWN
```
**After relocating non-NIC devices to high PCI slots (0x10+)** — hot-plugged NIC gets slot 0x04:
```
$ virsh qemu-agent-command i-2-1681-VM '{"execute":"guest-network-get-interfaces"}'
"name": "ens3", "hardware-address": "02:01:01:76:01:35"
"name": "ens4", "hardware-address": "02:aa:bb:cc:dd:01" ← sequential!
```
## Impact
- Hot-plugged NICs get `ens9`, `ens10`, etc. instead of sequential names
- Guest network configuration (cloud-init, CloudStack userdata) fails for the new NIC
- NIC shows as DOWN with no IP address in the guest
- Only a full VM reboot reassigns sequential PCI slots
## Fix
This PR adds explicit PCI slot assignment in `LibvirtPlugNicCommandWrapper`:
1. Query the domain XML to find all PCI slots currently in use
2. Find the highest PCI slot used by existing NICs
3. Assign the next free slot after the last NIC slot to the new NIC
**Note:** For full sequential naming, a follow-up change is needed to assign non-NIC PCI devices to higher slot numbers (e.g., 0x10+) during VM creation in `LibvirtComputingResource.createDevicesDef()`. This would leave slots 3-15 available for NICs, allowing hot-plugged NICs to always get sequential names.
## How to Reproduce
1. Create a KVM VM with 2 NICs
2. While running, add a 3rd NIC via CloudStack API/UI
3. Check inside guest: `ip link show` — the new NIC will be `ens9` (not `ens5`)
4. The NIC will be DOWN with no IP configuration
## Versions
- CloudStack 4.20
- QEMU 8.2, Libvirt 9.x
- Guest: Debian 12 (systemd-networkd with predictable interface naming)
Contributor guide
Research direction
Read LibvirtPlugNicCommandWrapper.java, especially execute(), vifDriver.plug(), and vm.attachDevice(), then compare the slot handling in LibvirtReplugNicCommandWrapper. Done means the hot-plug path assigns the next available PCI slot after existing NICs so the guest receives sequential interface naming; the issue also notes a follow-up in LibvirtComputingResource.createDevicesDef().
Written by the indexing model from the issue text.
Assessment
- Tech stack
- java
- Domain
- cloud, infrastructure
- Issue type
- Bug
- Difficulty
- 3/5
- Estimated time
- 1-2 days
- Activity status
- Stale
- Clarity
- Clearly specified
- Newbie friendliness
- 15/100