On Nov 19, 2004, at 14:18, Manfred Spraul wrote:

Hi,

I don't like the polling/interrupt setup part:
- for a nic driver, there is no irq line that could be requested by 
mii_phy_irq_enable().
- if the mii bus driver uses it's own timers, then locking within the 
nics will be more difficult.

I'm not sure I accept the argument that locking will be more difficult. 
  Jason's patch requires that a callback be registered for the interrupt 
or the polling (My update has a similar scheme).  The function you 
register is essentially like an extra interrupt, except it is never 
invoked at interrupt time.  All the function has to do is react 
properly to link state.  If, previously, you checked link state in your 
interrupt handler, you could still do it there, I suspect.

Could you make that part optional? For a nic driver, I would prefer if 
I could just call the ->startup part without the request_irq. If the 
nic irq handler notices that the nic got an event, then it would call 
an appropriate mii_bus function.

I think it would be doable to arrange the interface such that drivers 
could adopt only the PHY configuration infrastructure, and not any of 
the polling/interrupt infrastructure.  Of course, as it is, it is at 
least WHOLLY optional, so no driver has to use it at all.

This also applies for something like /dev/phy/xy: With natsemi, it 
would be very tricky to add proper locking. The nic as an internal phy 
and an external mii bus. The internal phy is partially visible on the 
external bus and any accesses to the phy id of the internal phy on the 
external bus cause lockups. No big deal, I just move the internal phy 
around [the phy id doesn't matter], but I would prefer if I have to do 
that just for ethtool, not for multiple interfaces.

I agree with this point -- Accessing the PHY through /dev registers is 
a recipe for some mess.  Though I could be convinced that it is 
manageable.  I do think, however, that the ethtool interface is 
sufficient to the task.