Hi!
Background
==========
This series adds a generic ethtool loopback framework with GET/SET
netlink commands, using a component/name/id/depth model that
represents loopback points across the network path.
This is v2.
Design
======
The loopback model uses four axes to identify a loopback point:
1. Component (MAC, PHY, MODULE) -- identifies the Linux driver
boundary that owns the loopback. This maps to the Linux driver
model, not to 802.3 sublayers directly.
2. Name -- identifies the sublayer within the component using IEEE
802.3 vocabulary (e.g. "mac", "pcs", "pma", "pmd", "mii" for MAC
and PHY components; "cmis-host", "cmis-media" for MODULE). The
name is a free-form string to avoid over-constraining the enum,
but recommended names are documented with references to 802.3
clauses.
3. Id -- optional instance selector within a component type (e.g.
PHY index from phy_link_topology, port number). Defaults to 0
when there is only one instance.
4. Depth -- ordering index within a component instance. When a
component has multiple loopback points of the same type (e.g. two
PCS blocks inside a rate-adaptation PHY), depth distinguishes
them. Lower values are closer to the host, higher toward the
line/media. Defaults to 0 when there is only one loopback point
per (component, name) tuple.
Direction uses IEEE 802.3 terminology: "local" (host TX -> looped back
-> host RX, traffic originating from host returns to host) and
"remote" (line RX -> looped back -> line TX, traffic from far end
returns to far end). Direction is defined from the component's own
viewpoint -- this convention holds regardless of where the component
sits in the system topology.
Cross-component ordering comes from existing infrastructure (PHY link
topology, phy_index). Multi-netdev shared-resource loopbacks (QSGMII,
port breakout) are out of scope; drivers should only expose loopbacks
they can scope to a single netdev.
A filtered DUMP (with a dev-index in the header) lists all loopback
entries for that netdev; an unfiltered DUMP iterates over all netdevs.
The doit handler supports exact lookup by (component, name, id).
V2 design rationale
===================
Why keep name as a string for now? The review discussed replacing name
with an enum. I decided not to do that in v2.
Pros:
- enforces more consistency across drivers
- makes userspace parsing simpler
- makes selftests stricter
- keeps the UAPI taxonomy explicit
Cons:
- freezes the stage taxonomy before we fully understand the MAC side
- risks forcing unlike hardware points into the same enum bucket
- makes later extension more expensive
- may standardize the spelling without standardizing the actual
semantics
This series still does not attempt to model the full datapath as a DAG
or expose a generic topology dump. That would likely be useful for
future diagnostics, traffic generators, and deeper validation, but it
also looks like a larger design problem than loopback itself.
The intent of v2 is not to solve that larger problem now, but also not
to block it:
- component gives the Linux ownership boundary
- name keeps room for stage-specific vocabulary
- id selects the component instance
- depth orders loopback points within that instance
If a richer topology API is needed later, it should be additive rather
than requiring this loopback API to be replaced.
Oleksij's more topology-heavy use cases are therefore not fully solved
by this series, but the current model should still be extendable for
them.
CMIS support
============
The Common Management Interface Specification (CMIS) defines four
diagnostic loopback types, characterized by location (Host or Media
Side) and signal direction:
- Host Side Input (Rx->Tx) -- local
- Host Side Output (Tx->Rx) -- remote
- Media Side Input (Rx->Tx) -- local
- Media Side Output (Tx->Rx) -- remote
Support is detected via Page 13h Byte 128, and loopback is controlled
via Page 13h Bytes 180-183 (one byte per type, one bit per lane).
The CMIS helpers work entirely over get/set_module_eeprom_by_page, so
any driver that already has EEPROM page access gets module loopback
without new ethtool_ops or driver changes.
Currently, only mellanox/mlxsw, and broadcom/bnxt support CMIS
operations. I'll follow-up with mlx5 support.
Implementation
==============
Patch 1/12 ethtool: Add dump_one_dev callback for per-device sub-iteration
Adds the dump_one_dev callback to ethnl_request_ops, the ifindex and
pos_sub fields to ethnl_dump_ctx, and the dispatch logic in
ethnl_default_start() and ethnl_default_dumpit(). No functional
change; no command uses dump_one_dev yet.
Patch 2/12 ethtool: Convert per-PHY commands to dump_one_dev
Converts PSE, PLCA, PHY, and MSE commands from the separate
ethnl_perphy_{start,dumpit,done} handlers to use the generic
dump_one_dev callback via a shared ethnl_perphy_dump_one_dev()
function.
Patch 3/12 ethtool: Add loopback netlink UAPI definitions
Adds the YAML spec and generated UAPI header for the new
LOOPBACK_GET/SET commands. Each loopback entry carries a component
type, optional id, name string, depth, supported directions
bitmask, and current direction.
Patch 4/12 ethtool: Add loopback GET/SET netlink implementation
Implements GET/SET dispatch in a new loopback.c. GET uses the
dump_one_dev infrastructure for dump enumeration (by flat index)
and supports doit exact lookup by (component, id, name) via
parse_request. SET switches on the component and calls the right
handler per entry. No components are wired yet.
Patch 5/12 ethtool: Add CMIS loopback helpers for module loopback control
Adds cmis_loopback.c with the MODULE component handlers and wires
them into loopback.c's dispatch. GET enumerates entries by index
(ethtool_cmis_get_loopback_by_index) or looks up by name
(ethtool_cmis_get_loopback). SET (ethtool_cmis_set_loopback)
resolves name to control byte indices and enforces mutual
exclusivity.
Patch 6/12 selftests: drv-net: Add loopback driver test
Adds loopback_drv.py with generic tests that work on any device
with module loopback support: enable/disable, direction switching,
idempotent enable, and rejection while interface is up.
Patch 7/12 ethtool: Add MAC loopback support via ethtool_ops
Extends struct ethtool_ops with three loopback callbacks for
driver-level MAC loopback: get_loopback (exact lookup by name/id),
get_loopback_by_index (dump enumeration), and set_loopback. Wires
the MAC component into loopback.c's dispatch. For dump enumeration,
MAC entries are tried first, then MODULE/CMIS entries follow at the
next index offset.
Patch 8/12 netdevsim: Add MAC loopback simulation
Implements the three ethtool loopback ops in netdevsim. Exposes a
single MAC loopback entry ("mac") with both local and remote
support. State is stored in memory and exposed via debugfs under
ethtool/mac_lb/{supported,direction}.
Patch 9/12 selftests: drv-net: Add MAC loopback netdevsim test
Adds loopback_nsim.py with netdevsim-specific tests for MAC
loopback: entry presence, SET/GET round-trip with debugfs
verification, and error paths.
Patch 10/12 MAINTAINERS: Add entry for ethtool loopback
Adds a MAINTAINERS entry for the ethtool loopback subsystem covering
the core loopback and CMIS loopback netlink implementation, and the
associated selftests.
Patch 11/12 netdevsim: Add module EEPROM simulation via debugfs
Adds get/set_module_eeprom_by_page to netdevsim, backed by a
256-page x 128-byte array exposed via debugfs.
Patch 12/12 selftests: drv-net: Add CMIS loopback netdevsim test
Extends loopback_nsim.py with netdevsim-specific tests that seed the
EEPROM via debugfs: capability reporting, EEPROM byte verification,
combined MAC + MODULE dump, and error paths.
Changes since v1
================
- Split dump_one_dev infrastructure patch into two: one adding the
generic callback infrastructure, one converting per-PHY commands
to use it. (Jakub)
- Used Jakub's suggested pattern for filtered dumps: initialize both
ifindex and pos_ifindex in _start(), then break early in the loop
when pos_ifindex diverges, avoiding a separate single-device code
path. (Jakub)
- Dropped PCS as a top-level component. Components now map to Linux
driver boundaries: MAC, PHY, MODULE. Sublayer granularity (pcs,
pma, pmd, mii) lives in the name attribute using IEEE 802.3
vocabulary. (Maxime)
- Renamed near-end/far-end to local/remote per IEEE 802.3
terminology. (Maxime)
- Added depth field to ETHTOOL_A_LOOPBACK_ENTRY for ordering
multiple loopback points within a single component instance (e.g.
rate-adaptation PHY with two PCS blocks). (Jakub)
- Documented the viewpoint convention: direction is always from the
component's own perspective, local = toward host, remote = toward
line, regardless of system topology. (Oleksij)
- Expanded component doc strings to explain that the name attribute
uses IEEE 802.3 sublayer vocabulary. (Andrew, Maxime)
Changes since RFC v2
====================
- Switched LOOPBACK_GET from doit-with-array to dumpit, where each
loopback entry is a separate netlink message. Uses the new generic
dump_one_dev sub-iterator infrastructure instead of duplicating the
perphy dump pattern. (Maxime)
- u32 to u8 to represent the enums in the YAML. (Maxime)
- Tried to document the YAML better. (Andrew)
- Added doit exact lookup by (component, id, name) via
parse_request, so single-entry GET doesn't need a flat index.
- Added MAC loopback support via three new ethtool_ops callbacks
(get_loopback(), get_loopback_by_index(), set_loopback()) with
netdevsim implementation and tests.
- Added MAINTAINERS entry.
Opens/limitations
=================
- mlx5 CMIS support is still not part of the series.
- PHY loopback is defined in the UAPI but not yet implemented.
- No per-lane support -- loopback is all-or-nothing (0xff/0x00)
across lanes.
Related work
============
[1] Generic loopback support, v1
https://lore.kernel.org/netdev/[email protected]/
[2] Generic loopback support, RFC v2
https://lore.kernel.org/netdev/[email protected]/
[3] New loopback modes
https://lore.kernel.org/netdev/[email protected]/
[4] PHY loopback
https://lore.kernel.org/netdev/[email protected]/
[5] bnxt_en: add .set_module_eeprom_by_page() support
https://lore.kernel.org/netdev/[email protected]/
[6] net/mlx5e: Implement set_module_eeprom_by_page ethtool callback
https://lore.kernel.org/netdev/[email protected]/
Björn Töpel (12):
ethtool: Add dump_one_dev callback for per-device sub-iteration
ethtool: Convert per-PHY commands to dump_one_dev
ethtool: Add loopback netlink UAPI definitions
ethtool: Add loopback GET/SET netlink implementation
ethtool: Add CMIS loopback helpers for module loopback control
selftests: drv-net: Add loopback driver test
ethtool: Add MAC loopback support via ethtool_ops
netdevsim: Add MAC loopback simulation
selftests: drv-net: Add MAC loopback netdevsim test
MAINTAINERS: Add entry for ethtool loopback
netdevsim: Add module EEPROM simulation via debugfs
selftests: drv-net: Add CMIS loopback netdevsim test
Documentation/netlink/specs/ethtool.yaml | 142 ++++++
MAINTAINERS | 6 +
drivers/net/netdevsim/ethtool.c | 147 +++++++
drivers/net/netdevsim/netdevsim.h | 15 +
include/linux/ethtool.h | 28 ++
.../uapi/linux/ethtool_netlink_generated.h | 67 +++
net/ethtool/Makefile | 2 +-
net/ethtool/cmis_loopback.c | 407 ++++++++++++++++++
net/ethtool/loopback.c | 351 +++++++++++++++
net/ethtool/mse.c | 1 +
net/ethtool/netlink.c | 284 ++++--------
net/ethtool/netlink.h | 49 +++
net/ethtool/phy.c | 1 +
net/ethtool/plca.c | 2 +
net/ethtool/pse-pd.c | 1 +
.../testing/selftests/drivers/net/hw/Makefile | 2 +
.../selftests/drivers/net/hw/loopback_drv.py | 227 ++++++++++
.../selftests/drivers/net/hw/loopback_nsim.py | 343 +++++++++++++++
18 files changed, 1865 insertions(+), 210 deletions(-)
create mode 100644 net/ethtool/cmis_loopback.c
create mode 100644 net/ethtool/loopback.c
create mode 100755 tools/testing/selftests/drivers/net/hw/loopback_drv.py
create mode 100755 tools/testing/selftests/drivers/net/hw/loopback_nsim.py
base-commit: d1e59a46973719e458bec78d00dd767d7a7ba71f
--
2.53.0
