[PATCH net v4 00/10] net: au1000: Bug fixes

From: Florian Fainelli

Date: Fri Nov 20 2026 - 19:48:31 EST


This series fixes RX descriptor recycling, receive filtering, MDIO error
reporting, TX ring synchronization, MAC/PHY lifecycle ordering, and TX
bounce-buffer bounds in the Au1xxx Ethernet driver. It also adds the
LAN83C183 PHY resume delay needed for MII clock recovery on the MTX-1.

Changes in v4:
- Correct the PHY-disconnect commit message: the attachment path uses
polling, so describe device/module reference leaks and the stale
netdev attachment instead of surviving PHY interrupts. Document the
existing repeated-reset hazard after register_netdev() failure.
- Narrow the TX critical sections to descriptor reservation and final
publication. Copy and pad the packet outside the IRQ-disabled lock.
Clear TX_T_DONE and advance tx_tail before dropping the lock when
submission reclaims a completed descriptor, preventing double reclaim
by a completion interrupt during the unlocked copy.
- Document that the SMSC PHY driver must bind before MAC attachment.
The tested early-boot configuration uses CONFIG_SMSC_PHY=y; loading
smsc.ko after genphy has attached does not replace that binding.
- Add the LSI/SEEQ OUI/model comment and document that the LAN83C183 PHY
ID is shared with the L80227, which receives the same conservative
resume delay. L80227 hardware remains untested.
- Add MAC-managed system-sleep callbacks. Explicitly stop and quiesce
MAC RX/TX before stopping the PHY, including PHY states without a
link-down callback. Restore the PHY before MAC initialization on
resume. Handle inactive interfaces and keep failed resumes detached
with the MAC reset; prevent IRQ-disable nesting on a subsequent retry.
- Move the link-down quiesce wait outside the device lock and use a
sleepable wait with local interrupts restored. Retain existing atomic
waits in the duplex-change and reset paths.
- Extend validation with TX/PM sanitizer harnesses, simultaneous TX/RX
traffic, RX-stop descriptor observations, and device suspend/resume
tests on MTX-1.

Changes in v3:
- Drop the no-platform-data MAC-address patch: in-tree platform devices
always supply platform data, so this is defensive hardening rather
than a fix for an in-tree configuration.
- Drop the buffer descriptor free-list insertion patch: the affected
list is only rebuilt during teardown and is never consumed again.
The series now contains 10 patches instead of 12.
- Distinguish MDIO write-completion timeouts from pre-write busy
timeouts in the diagnostic message.
- Clarify that the LAN83C183 recovery-delay patch and the following
au1000 MAC/PHY sequencing patch are both required to fix MTX-1 reopen.
Keep the PHY patch first so its delay is available to the MAC fix.
- Remove the unreachable initialization-error branch and PHY rollback
from au1000_open(). Correct the commit message to distinguish the new
synchronous PHY resume from the existing initialization-before-
phy_start() ordering, and reconcile the reset comments.
- Expand the TX bounds-check commit message to cover plain AF_PACKET
sendmsg(), PACKET_TX_RING, and oversized software-GSO segments.
- Rebuild the kernel and root filesystem and repeat MTX-1 regression
testing after the review changes.

Changes in v2:
- Rebase onto net.
- Reset the MAC after PHY/MDIO teardown, and avoid MAC control accesses
when its clock is already disabled.
- Use sleepable polling for MDIO write completion.
- Advance tx_tail when submission reclaims a completed TX descriptor,
and consume transmitted skbs outside the lock with an IRQ-safe helper.
- Restore receive filtering before enabling RX, with the address-list
and device locks held to keep the programmed filter consistent.
- Resume the PHY synchronously before MAC initialization, clear cached
link state on open, and use acknowledged link/duplex state under the
device lock. Place the LAN83C183 recovery-delay patch before this fix.
- Add a TX length check to prevent overflowing the DMA bounce buffer.

Testing:
The current series was rebuilt with Buildroot and tested on a Meshcube
MTX-1, using the Alchemy PROM Ethernet-address lifetime fix as a separate
boot prerequisite. That prerequisite is not included in this series.
CONFIG_MIPS_AU1X00_ENET=y and CONFIG_SMSC_PHY=y were used, and boot logs
confirmed attachment to the SMSC LAN83C183 driver.

The final run passed all 11 boot-smoke checks, including 15 down/up
cycles across normal, promiscuous and all-multicast modes, opened/closed
and never-opened removal/rebind, and connectivity recovery. Five
additional device suspend/resume cycles passed: three with the
interface up, one down, and one never opened after rebind. These used
pm_test=devices with freeze and CONFIG_PM_DEBUG=y. The final ping
received all 10 replies, and the kernel-log check reported no findings.
Actual platform suspend-to-RAM was not tested.

An earlier run with the revised TX locking completed 20 seconds of
bidirectional iperf3 traffic at approximately 14.4 Mbit/s TX and
40.9 Mbit/s RX. RX missed-frame errors increased by 10, while TX errors
did not increase. There was no before/after overrun or IRQ-latency
measurement; this test does not establish elimination of RX overruns.

The extracted TX functions passed 432 ring, copy, padding, IRQ-state
and completion-interleaving cases under ASan/UBSan, plus busy and
oversized-packet paths. Tests rejected the original locked copy and a
variant omitting the early TX_T_DONE clear. A separate PM harness
checked up/down and PHY-less cases, missing link-down callbacks,
error cleanup and retry, and the unlocked link-down wait.

A logging-only kernel captured 44 RX descriptor snapshots across 11
stops during duplex changes under incoming traffic. Nine descriptors
had advanced addresses with EN set and DN clear before stopping; all
nine had DN set after the existing 10 ms wait. The reported state with
an advanced address and DN clear after stopping was not reproduced.
This finite MTX-1 test does not establish behavior for every silicon
revision or link-loss timing; no RX reload change was made for that path.

Full AF_PACKET overflow triggers and hardware fault injection were not
exercised.

Sashiko AI review identified issues addressed in the revisions. OpenAI
Codex assisted source and datasheet analysis, code and changelog
revisions, patch consolidation, test harnesses and regression testing.

Florian Fainelli (10):
net: au1000: recycle RX descriptors on allocation failure
net: au1000: clear all-multicast mode when restoring hash filtering
net: au1000: disconnect the PHY before freeing the MDIO bus
net: au1000: report MDIO transaction timeouts
net: au1000: serialize TX completion with packet submission
net: au1000: restore receive filtering before enabling RX
net: au1000: restore RX buffer addresses when restarting the MAC
net: phy: smsc: wait for LAN83C183 power-down recovery
net: au1000: keep the MAC stopped until the PHY link is ready
net: au1000: reject frames larger than the TX bounce buffer

drivers/net/ethernet/amd/au1000_eth.c | 248 +++++++++++++++++++++-----
drivers/net/phy/Kconfig | 2 +-
drivers/net/phy/smsc.c | 27 +++
3 files changed, 233 insertions(+), 44 deletions(-)


base-commit: 145480b28ca82e9fae789427cc4569e596515a32
--
2.53.0