[PATCH net-next v5 0/4] net: mdio: Clause 45 support for mdio-bcm-unimac
From: Florian Fainelli
Date: Fri Nov 20 2026 - 14:44:00 EST
This patch series adds support for Clause 45 MDIO operations within the
mdio-bcm-unimac driver. The series breaks things up in this way:
- make the C45 bus scan done by the MDIO core carry on when nothing
answers at an address, as the C22 scan already does
- scope the existing defines to make it clear what is C22 vs. C45
specific
- rework the polling completion loop to be a tighter loop
- add support for clause 45 operations
Changes in v5:
- patch 1:
- the commit message now says that mdio-bcm-unimac only provides
read_c45() with patch 4, which enables the UniMAC case, so that a
standalone backport knows about that dependency
- the commit message now lists the callers other than the bus scan that
see -ENODEV instead of -EIO (of_mdio, sfp, mxl-gpy), which is what C22
callers already get
- the vendor MMD presence check done while reading the identifiers and the
reads of the per-MMD identifiers are intentionally not mapped, since
they only happen once a device has answered; say so in the commit
message and reword the kernel-doc of get_phy_c45_ids() and
phy_get_c45_ids() to document "a negative error code" rather than -EIO
as the only alternative to -ENODEV
- patches 2 and 3: no changes
- patch 4:
- check that the controller is idle after every cycle (the C45 ADDR and
data cycles as well as the C22 ones) instead of only after the ADDR
cycle, since a wait_func() such as bcmgenet_mii_wait() returns 0 even
when it timed out with START_BUSY still set, which could return the
command bits instead of the read data or switch MDIO_CFG in the middle
of a frame
- the commit message says that unimac_mdio_resume() waiting for the bus
lock means that resume may wait for an in-flight MDIO access
- the commit message says that the C45 scan is slower than the C22 one on
the platform device GENET registers without a Device Tree node (e.g.:
ACPI) because bcmgenet_mii_wait() sleeps per frame until GENET's
interrupt is registered. This is not worked around in this series;
making bcmgenet_mii_wait() poll in that window would be a separate
GENET change
Changes in v4:
- patch 1:
- when the read of the devices in package registers of an MMD fails with
-EIO or -ENODEV, treat it like an all Fs reply and probe MMD 0 instead
of returning -ENODEV right away, so that PHYs that only respond on
MMD 0 (e.g.: Cortina CS4315/CS4340) are still found on controllers
that report a bus error for an MMD that does not answer. The remaining
MMDs are not probed, to keep the cost of scanning an empty address
bounded. -ENODEV is only returned if MMD 0 does not answer either
- add a Fixes: tag
- update the kernel-doc of get_phy_c45_ids() and phy_get_c45_ids()
- the code changed since Nicolai's Reviewed-by tag was given, which was
kept, please let me know if it should be dropped
- patch 2: no changes
- patch 3: reword the comments about the poll timeout, which applies to each
frame, a C45 access waits once for each of its two frames
- patch 4: no code changes, the commit message now mentions that the C45
scan also runs for the platform device GENET registers without a
Device Tree node (e.g.: ACPI), since it does not mask any address
Changes in v3:
- patch 1: collect Reviewed-by tag from Nicolai Buchwitz, no code changes
- patches 2 and 3: no changes
- patch 4:
- after the ADDR cycle, make sure the controller is idle before starting
the READ/WRITE cycle, since a wait_func() such as bcmgenet_mii_wait()
returns 0 even when it timed out with START_BUSY still set
- take bus->mdio_lock in unimac_mdio_resume() around
unimac_mdio_clk_set(), such that its read-modify-write of MDIO_CFG
cannot race with the C22/C45 clause switch done by the accessors
Changes in v2:
- add patch 1, since providing read_c45/write_c45 makes the MDIO core scan
for C45 PHYs on all addresses that are not masked. On a bus registered
without a device tree node (e.g.: GENET with ACPI) a read from an empty
address fails with -EIO, which get_phy_c45_ids() passed up instead of
mapping it to -ENODEV, and made the whole bus registration fail. Reported
by Nicolai Buchwitz.
- patches 2 and 3: collect Reviewed-by tags from Andrew Lunn and Nicolai
Buchwitz, no code changes
- patch 4: no code changes
Florian Fainelli (4):
net: phy: Map C45 probe read errors to -ENODEV during bus scan
net: mdio: mdio-bcm-unimac: Define MDIO_OP_SHIFT
net: mdio: mdio-bcm-unimac: Poll for completion at 1us granularity
net: mdio: mdio-bcm-unimac: Add Clause 45 support
drivers/net/mdio/mdio-bcm-unimac.c | 162 +++++++++++++++++++++++++----
drivers/net/phy/phy_device.c | 50 +++++++--
2 files changed, 184 insertions(+), 28 deletions(-)
--
2.34.1