Re: pvops netconsole

[Jeremy Fitzhardinge <jeremy@goop.org>]

On 02/13/2010 08:56 PM, Brad Plant wrote:
> Actually, I tried something like this first because I first looked at pcnet32.c and e100.c:
>
> disable_irq(dev->irq);
> xennet_interrupt(dev->irq, dev);
> enable_irq(dev->irq);
>
> The same WARN_ONCE() was triggered.
>    

Given that netconsole works with e1000 (I presume; haven't tested it 
lately), I guess it needs a close look to see what differs from the 
netfront case.

>>> 3. Initially netconsole would not initialise because the interface didn't yet have a mac address as setup_netfront hadn't yet been called. I changed the module_init() in netconsole.c to late_initcall() if !CONFIG_MODULES&&   CONFIG_XEN. Is this allowed and sane?
>>>        
>> Probably not.  In practice every kernel will have modules and Xen
>> enabled, and it will change netconsole for all users.  I guess we
>> probably need to find some way to init the MAC address earlier, or
>> trigger netconsole init once the mac address is known.
>>      
> I figured it was dodgy. How would you init the MAC address earlier though? I've observed the xen devices being initialised at slightly different times with respect to other devices which suggests to me that they (and possibly all devices) are being initialised in separate threads. It would mean setting the MAC address before device initialisation yes?
>
> Not sure how correct this is either, but another solution might be to set the MAC address to say 00:00:00:00:00:01 when the netdev structure is first created. This lets netconsole initialise since the MAC address isn't all zeros or all 0xFF. By the time any net console logs are transmitted, setup_netfront will have already been called and the correct MAC address set so no packets will actually be sent with the above MAC address. I have attached a patch that does this.
>    

That seems a bit sketchy.

netfront is like a lot of network devices; it needs to query some 
external source to find its own mac address.  Mostly this means querying 
some eeprom over i2c or something, but in our case it means talking to 
xenstore to find the configured mac address.  I think the problem here 
is that this happens asynchronously, rather than synchronously in the 
net device's init routine.  One fix might be to synchronously wait for 
the mac address to be set before returning from the init function.

     J