Re: [RFC PATCH net-next 2/2] net: dsa: mt7530: introduce MMIO driver for MT7988 SoC

[Vladimir Oltean <olteanv@gmail.com>]

On Tue, Mar 28, 2023 at 04:00:30PM +0100, Daniel Golle wrote:
> On Tue, Mar 28, 2023 at 05:16:28PM +0300, Vladimir Oltean wrote:
> > On Tue, Mar 28, 2023 at 02:08:18PM +0100, Daniel Golle wrote:
> > > I agree that using regmap would be better and I have evaluated that
> > > approach as well. As regmap doesn't allow lock-skipping and mt7530.c is
> > > much more complex than xrs700x in the way indirect access to its MDIO bus
> > > and interrupts work, using regmap accessors for everything would not be
> > > trivial.
> > > 
> > > So here we can of course use regmap_read_poll_timeout and a bunch of
> > > readmap_write operations. However, each of them will individually acquire
> > > and release the mdio bus mutex while the current code acquires the lock
> > > at the top of the function and then uses unlocked operations.
> > > regmap currently doesn't offer any way to skip the locking and/or perform
> > > locking manually. regmap_read, regmap_write, regmap_update_bits, ... always
> > > acquire and release the lock on each operation.
> > 
> > What does struct regmap_config :: disable_locking do?
> 
> I thought I can't use that on a per-operation base because the
> instance of struct regmap_config itself isn't protected by any lock
> and hence setting disable_locking=false before calling one of the
> accessor functions may affect also other congruent calls to the
> accessors which will then ignore locking and screw things up.
> Please correct me if I'm wrong there.

It's not supposed to be used like that. You set disable_locking = true
once, and take care of locking from the mt7530 driver. Not before every
operation.

> Yet another way I thought about now could also be to have two regmap
> instances, one for locked and one for unlocked accessed to the same
> regmap_bus.

That would also be theoretically possible, but the above would be
equivalent and would require a single regmap.