> So this does not really solve my problem - even if we do get the lock
> by calling select(), we will simply release it as soon as our
> i2c_transfer() is finished (and will let the other master take it,
> breaking atomicity of the two operations I need to perform on target
> device) and this is exactly what I need to avoid.

I was assuming that once you have the lock from the arbitrator you can
be sure that the other master is not active and, thus, you maybe don't
need to read the status register at all? But I am probably wrong here.