Re: [PATCH 1/4] net: phy: balance disable/enable irq on change

[Jean-Michel Hautbois <jhautbois@gmail.com>]

2010/12/23 David Miller <davem@davemloft.net>:
> From: Jean-Michel Hautbois <jhautbois@gmail.com>
> Date: Sun, 19 Dec 2010 11:58:48 +0100
>
>> When phy interface changes its status, it calls phy_change() function.
>> This function calls the interrupt disabling functions for the driver
>> registered, but if this driver doesn't implement it, there is no IRQ
>> disabling. After doing the work, we call enable_irq and not the
>> respective driver function. This fixes it, as it could lead to an
>> unbalanced IRQ. Error code changed to EOPNOTSUPP.
>>
>> Signed-off-by: Jean-Michel Hautbois <jhautbois@gmail.com>
>
> This is completely bogus.
>
> First of all, there are 5 call sites for phy_change_interrupt() but
> you've only implemented the new semantics for two of those.
>
> Therefore, if we even wanted this, we should implement the behavior in
> phy_change_interrupt() itself instead of duplicating the logic at
> each and every call site.
>
> But we don't want this.

OK, I understand that point.

> It's not appropriate at all.  If a device lacks a way to turn
> interrupt off and on, using disable_irq() and enable_irq() is not
> necessarily correct.
>
> If the interrupt line is shared, for example, this will break
> everything.
>

OK, well, maybe is there at least one thing we could do : in
phy_change, instead of calling phy_disable_interrupts(), balanced by
enable_irq, we probably should use phy_enable_interrupts().