Re: [PATCH net-next 06/12] net: dsa: rzn1-a5psw: add Renesas RZ/N1 advanced 5 port switch driver

[Vladimir Oltean]

On Fri, Apr 15, 2022 at 12:23:30PM +0100, Russell King (Oracle) wrote:
> If ->shutdown has been called, the system is going down, and userspace
> is probably already dead.

There isn't anything preventing ->remove from being called after ->shutdown
has been called.

It all starts with the pattern that some driver authors prefer, which is
to redirect their ->shutdown method to ->remove. They argue that it
provides for a well-tested common path, so in turn, this pattern is
quite widespread and I'm not one to argue for removing it.

When such driver (redirecting ->shutdown to ->remove) is a bus driver
(SPI, I2C, lately even the fsl-mc bus), the implication is that the
controller will be unregistered on shutdown. To unregister a bus, you
need to unregister all devices on the bus too.

Due to implicit device ordering on the dpm_list, the ->shutdown() method
of children on said bus has already executed, now we're in the context
of the ->shutdown() procedure of the bus driver itself.

You can argue "hey, that's SPI/I2C and this is a platform driver, there
isn't any bus that unregisters on shutdown here", and you may have a
point there. But platform devices aren't just memory-mapped devices,
they can also be children of mfd devices on SPI/I2C buses. So in theory
you can see this pattern happen on platform devices as well.

This is the reason why I insist for uniformity in the DSA layer in the
way that shutdown is handled. People copy and paste code a lot, and by
leaving them with less variance in the code that they copy, subtle
differences that are not understood but do matter are less likely to
creep in.