This changes large master transfers to do shorter read/write rather than
return -EINVAL. User space will now be able to optimistically request a
large transfer and get at least some data.
This also removes comments suggesting on how to implement large
transfers. Current vme_master_* read and write implementations use CPU
copies that don't produce burst PCI accesses and subsequently no block
transfer on VME bus. In the end overall performance is quiet low and it
can't be fixed by doing direct copy to user space. Much easier solution
would be to just reuse kernel buffer.