docs: design notes for a Redfish/IPMI BMC driver

No code yet. The README explains why one Redfish driver with vendor quirk
modules beats a Supermicro-only IPMI driver, and keeps IPMI over LAN as a
read-only fallback. docs/ holds a draft for a new ManagementController
role base in napalm-device-types, the getter-to-resource map, the first
test device (Supermicro X10DRL-i, Redfish firmware 3.91), the security
points (IPMI RAKP hash disclosure, power control, credentials) and the
test plan with DMTF mockups and fixtures to record.
This commit is contained in:
Christian Manivong committed 2026-10-13 01:15:57 +02:00
commit 5f8a4ef975
9 files changed
+557

No files matched your search

+70
View File
@@ -0,0 +1,70 @@
# Role base draft: `ManagementControllerDriver`
Draft for napalm-device-types. Nothing here is decided; the decisions are tracked in the
netOrk options issue.
## Why a new role
The existing role bases (`SwitchDriver`, `FirewallDriver`, `AccessPointDriver`,
`HypervisorDriver`, `StorageDriver`, `OSDriver`, `PhoneDriver`, `MediaDriver`,
`ResidentialGatewayDriver`) each describe what a device *is*. A BMC is a separate computer
with its own network port, IP and MAC that manages *another* machine: it reads that
machine's sensors and switches its power, whatever OS the machine runs. None of the
existing contracts fits, and borrowing one (e.g. `OSDriver`) would make netOrk offer
package updates and service control for a BMC.
## Sketch
```python
class ManagementControllerDriver(DeviceTypeDriver):
"""Contract, not code."""
ROLE: str = "management_controller" # key exposed as device_class
TYPE_LABEL: str = "Management Controller"
USES_SSH: bool = False # Redfish is HTTPS; IPMI is UDP
if TYPE_CHECKING:
# The managed machine
def get_power_state(self) -> PowerStateDict: ...
def set_power_state(self, action: PowerAction) -> None: ... # On, GracefulShutdown,
# ForceOff, GracefulRestart,
# ForceRestart, PowerCycle
def get_hardware_inventory(self) -> HardwareInventoryDict: ... # CPUs, DIMMs, drives,
# PSUs, NICs, serials
def get_event_log(self, limit: int = 500) -> List[EventLogEntryDict]: ... # SEL
# The controller itself
def get_firmware_inventory(self) -> List[FirmwareDict]: ... # BMC, BIOS, others
def get_management_protocols(self) -> ManagementProtocolsDict: ... # IPMI, SSH,
# HTTP(S), SNMP,
# KVM, virtual media
```
Points to settle in napalm-device-types:
- **`reboot_host` means the BMC.** `HostRebootMixin` says a driver restarts its *own*
host, never one it manages ("a hypervisor restarts the hypervisor, not a VM"). For a BMC
that is `Manager.Reset`; the managed server goes through `set_power_state`. A driver for
a BMC must not map `reboot_host` to the server.
- **Sensors.** NAPALM's `get_environment` already has fans, temperatures and power supplies.
It also requires CPU and memory, which a BMC does not know for the host OS. Either return
`get_environment` with those empty, or add a `get_sensors` that also carries voltages and
power draw. Prefer reusing `get_environment` if the empty fields are acceptable.
- **Health metrics.** `HealthMetricsMixin` is SNMP/UCD-MIB. A BMC's health comes from Redfish,
so this role does **not** mix it in. If netOrk needs `get_health_metrics` for BMCs, the
driver implements it from Redfish, as `SwitchDriver` drivers do from their own sources.
- **`CommandChannelMixin`** comes with `DeviceTypeDriver` but stays unimplemented: no shell.
- **Write capability by `hasattr`.** `set_power_state` exists only on drivers that can do it,
so the IPMI fallback (read-only) simply does not define it.
- **Models.** `PowerStateDict`, `HardwareInventoryDict`, `EventLogEntryDict`,
`FirmwareDict`, `ManagementProtocolsDict` go into `models.py`.
- **Fingerprinting.** An unauthenticated `GET /redfish/v1/` returns the service root
(`RedfishVersion`, `UUID`, `Vendor` on newer services) on every Redfish BMC. That is a
vendor-neutral `PORT_SPECS`/`HTTP_FINGERPRINT` rule and can replace the hand-written
`idrac`/`ilo` platform strings in netOrk's discovery.
## Relation to the managed server
A BMC is its own device in netOrk (own IP, own credentials, own firmware and CVEs), linked
to the host it manages. How to link them (matching the server's serial or system UUID from
`Systems/{id}` against the host's DMI data, or by hand) is a netOrk decision, not part of
this contract. The driver exposes the identifiers needed for matching:
`SerialNumber`, `UUID` and the host's MAC addresses from `Systems/{id}/EthernetInterfaces`.