phy_attach_direct() binds the generic driver by hand, and the probe it calls is phy_probe(), which replaces phydev->irq with PHY_POLL before either point the hand-bind can fail at. That failure unwinds on a label of its own, which does not go through phy_detach(), so the substitution outlives a bind cycle that never completed and a later attach finds a PHY that can only be polled. Found on a Keenetic KN-1012 while placing the restore of the previous patch, as the other exit of the same bind cycle. Save phydev->irq on entry and put it back on that label. The unwind runs inside the call that made the substitution, so the value from before it is known exactly. The bus table the previous patch reads from would be wrong here twice over: it does not hold a PHY_MAC_INTERRUPT that a MAC wrote into phydev->irq alone, and the label is also reached when a second attach of an attached PHY fails its bind, where the field is live. The store is not ordered against a concurrent bind: this unwind, like the hand-bind it undoes, runs without the device lock that device_bind_driver() asks its callers to hold. Fixes: 6d9f66ac7fec ("net: phy: Fix PHY module checks and NULL deref in phy_attach_direct()") Assisted-by: LLM Signed-off-by: Aleksei Sviridkin --- Notes: Both points the hand-bind can fail at are reachable. phy_probe() reaches genphy_read_abilities() through genphy_driver's .get_features, and that returns the error from phy_read(phydev, MII_BMSR); device_bind_driver() returns whatever driver_sysfs_add() got, from either of its two sysfs_create_link() calls or from the coredump attribute. A failed genphy bind leaves the device with no driver bound at all, so the next driver to arrive binds directly and never goes through phy_detach(). That is why patch 3 cannot cover this path, and why the Fixes: tag here is 6d9f66ac7fec rather than the one patch 3 carries. That commit did not introduce the lost number - the substitution is far older - it created this second exit from the bind cycle, splitting the failure off the label that calls phy_detach(). Before it, patch 3 alone would have covered this, so that is where the backport range for this one starts. Exercised on the board described in patch 3, with a debug-only module parameter that fails the hand-bound generic probe once for one MDIO address. The connect then ends in -EIO rather than the -EINVAL of the validation path, so the unwind takes the label this patch touches. Measured again for this version, since the value now comes from the local: two images of the distribution's 6.18.52 kernel differing only by this patch, injected failure at 2.0 s, real driver bound at 6.4 s. phydev->irq afterwards reads -1 with patch 3 alone and 15 with this one; the three switch ports read 79, 80 and 81 in both. One difference between the injector and a real failure, since it does not affect what was measured but should not be implied away: a genuine error inside phy_probe() leaves through its out: label, which re-asserts the PHY reset before returning, while the injector returns earlier than that. Neither path touches phydev->irq. drivers/net/phy/phy_device.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/drivers/net/phy/phy_device.c b/drivers/net/phy/phy_device.c index a9c71a286118..8bfb154402ad 100644 --- a/drivers/net/phy/phy_device.c +++ b/drivers/net/phy/phy_device.c @@ -1755,6 +1755,7 @@ int phy_attach_direct(struct net_device *dev, struct phy_device *phydev, struct mii_bus *bus = phydev->mdio.bus; struct device *d = &phydev->mdio.dev; struct module *ndev_owner = NULL; + int irq = phydev->irq; int err; /* For Ethernet device drivers that register their own MDIO bus, we @@ -1896,6 +1897,7 @@ int phy_attach_direct(struct net_device *dev, struct phy_device *phydev, error_module_put: module_put(d->driver->owner); + phydev->irq = irq; phydev->is_genphy_driven = 0; d->driver = NULL; error_put_device: -- 2.53.0