John Newbigin wrote:
>
> I have found a solution to another problem with the bonding driver.
>
> If you ifconfig down an enslaved eth device, the bonding driver is not
> notified and if there is an attempt to use the device then a crash
> results. I have added a one line fix to dev_close in net/core/dev.c
>
> /* if the device is a slave we should not touch it*/
> if(dev->flags&IFF_SLAVE)
> return -EBUSY;
>
> I put it after the check to see if the interface is up.
>
> This may not be the best solution but it prevents the kernel crashes and
> as bond_release is not implemented there is not much point making the
> device release it's self.
>
> Since I made this change and the one in my last message I have not been
> able to produce a crash.
>
> (here is a script which will cause the crash on 2.2.16)
> #!/bin/bash
> insmod eepro100
> insmod bonding
> ifconfig bond0 192.168.1.2 up
> ifenslave bond0 eth0
> ifconfig eth0 down
> ifconfig bond0 down
> ifconfig bond0 192.168.1.2 up
> ifconfig bond0 down
>
> Again, if someone could confirm this fix it would be good.
>
> John.
I agree.
I've looked at the other users of IFF_SLAVE (eql.c and sbni.c). They appear
to be OK with this change. In fact I wonder if they're vulnerable to the
same problem in kernel 2.2.
|