[PATCH net-next v6 0/3] net: phylink: wait for a PHY that probes after the MAC

[Aleksei Sviridkin <f@lex.la>]

On the Keenetic KN-1012 (MT7981B with an MT7531 switch), the Airoha
EN8811H behind lan4 has its PHY driver built as a module on the root
filesystem. The switch sets up its ports before that filesystem is
mounted, so the port is validated against the generic driver, fails its
phy-mode and stays dead for the uptime. DSA does not retry it.

Patch 1 lets the PHY node say so with needs-host-firmware. Patch 2 makes
phylink poll for such a PHY instead of giving up, for a MAC that opts
in. Patch 3 opts in DSA user ports of switch drivers that set a flag,
and sets it in mt7530.

The poller can still lose a race against an unbind of the PHY driver,
between its readiness check and the attach. That window is phylib's:
any phy_attach_direct() caller racing an unbind has it. The attach
guard series [1] closes it.

Tested on that board with the series backported to its OpenWrt 6.18
kernel, together with a940003f44e7 and 07d995873960 (the mt7530
.get_stats64 atomic-context fix). The kernel had lockdep and
DEBUG_ATOMIC_SLEEP on. Local debug parameters, not part of the series,
drove the error paths. They fail the connect after a successful attach,
ignore the opt-in, fail the generic attach of a PHY with no driver, and
add a sleep after the switch shutdown.

 - boot: the PHY driver loaded its firmware at 8.7 s. lan4 attached at
   26.3 s, when the port was brought up. Link up at 1 Gb/s.
 - port kept down for 10 s with the PHY driver bound: no poll and no
   attach. "ip link set lan4 up" attached it a second later.
 - opt-in ignored: the old behaviour, the generic driver took the PHY.
 - two injected failures: "failed to connect late PHY: -EIO" twice, a
   second apart, and the third attempt attached. Link up at 1 Gb/s.
 - failures that do not stop: four attempts and one "giving up". No
   poll after that, also not after a down/up.
 - four failed generic attaches with no driver bound: nothing logged as
   a failure and no retry spent. The port attached once the driver
   bound.
 - switch unbound while the poller waited: nothing oopsed.
 - reboot with a 3 s sleep after the switch shutdown: the poll ran
   every second up to the shutdown and not once in the sleep. On v5 the
   same test polled 3 times in the sleep.

No in-tree device tree sets needs-host-firmware yet. The board is
supported out of tree, in OpenWrt.

Changes in v6 (since v5):
https://lore.kernel.org/r/20261001130208.105558-1-f@lex.la/
 - Defer only in PHY mode and without an SFP cage. In-band, the PCS
   could bring the carrier up with no PHY attached. An SFP PHY could
   take the port while the wait is armed, and its removal dropped the
   wait.
 - Poll only while phylink is started. A port that is down attaches at
   its next up, and nothing touches the PHY after a switch shutdown.
 - A failed attach after the PHY driver went away no longer counts as a
   failed connect.
 - The patch 2 message no longer says that MACs connecting from
   ndo_open recover on the next open. They do not when the PHY driver
   loads while the generic one holds the PHY.
 - The dsa.h comment states the rule: .port_enable must not use its phy
   argument.

[1] https://lore.kernel.org/r/20261001130120.104628-1-f@lex.la/

Aleksei Sviridkin (3):
  dt-bindings: net: ethernet-phy: add needs-host-firmware
  net: phylink: wait for PHYs that are known to probe late
  net: dsa: let user ports wait for a PHY that probes late

 .../devicetree/bindings/net/ethernet-phy.yaml |   6 +
 drivers/net/dsa/mt7530.c                      |   1 +
 drivers/net/phy/phylink.c                     | 232 +++++++++++++++++-
 include/linux/phylink.h                       |   5 +
 include/net/dsa.h                             |   5 +
 net/dsa/user.c                                |   1 +
 6 files changed, 242 insertions(+), 8 deletions(-)


base-commit: 47a1446725732cd3996edf607e8739334bbf4d78
-- 
2.53.0