net: dsa: mt7530: serialize the regmap IRQ chip like every other user

The switch register regmap is created with .disable_locking = true;
every other user in this driver calls mt7530_mutex_lock()/unlock()
around it, which takes priv->bus->mdio_lock, since the underlying
mt7530_regmap_read()/write() issue raw, unserialized bus->read()/
write() MDIO transactions.

mt7530_setup_irq() hands this same unlocked regmap straight to
devm_regmap_add_irq_chip_fwnode(), whose threaded IRQ handler then
calls regmap_read()/regmap_update_bits() on it without ever calling
mt7530_mutex_lock(). An interrupt firing while another thread is
mid-transaction on the same regmap (e.g. a paged register access, or
an indirect PHY access) can interleave with the IRQ handler's own
paged access and corrupt page selection on either side.

Use struct regmap_irq_chip's handle_mask_sync hook to call
mt7530_mutex_lock()/unlock() around the mask register write regmap-irq
issues whenever a consumer of one of the mapped sub-IRQs enables,
disables, requests or frees its line. This needs a per-device copy of
mt7530_regmap_irq_chip, since devm_regmap_add_irq_chip_fwnode() keeps
a pointer to it rather than copying it.

handle_pre_irq/handle_post_irq, which would additionally cover the
status read and ack write the threaded handler does directly, bracket
the whole handler including its handle_nested_irq() calls. Lockdep
caught this on hardware: those calls reach phy_interrupt() for the
per-port PHY IRQ lines mapped through this chip, which takes
phydev->lock, while phy_attach_direct() and this driver's own indirect
PHY access already establish the opposite order (phydev->lock, then
priv->bus->mdio_lock) elsewhere. Using them here would close that
cycle, so they are not used.

regmap_irq_sync_unlock() also has its own init_ack_masked path, used
by this chip, which unconditionally does its own regmap_write() to ack
currently-masked IRQs; that path has no per-driver hook. Together with
the threaded handler's own status read and ack write, these stay
unprotected -- a narrower, harder-to-hit gap than the recurring mask
sync above -- and will be closed once the switch regmap moves to
regmap's own locking in the driver-wide register access cleanup.

Signed-off-by: Daniel Golle <daniel@makrotopia.org>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://patch.msgid.link/818840879e9cd20f8d568789da29b3474c8f3ab9.1785811140.git.daniel@makrotopia.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

0 files changed, 0 insertions, 0 deletions