Attached are two patches. First one implements new function
mca_pml_ob1_send_requst_copy_in_out(req, offset, len) that sends
given range of the request by copying data in/out internal buffers.
It also changes the behaviour of the pipeline protocol to send data from
an end of a user buffer. The second patch fixes deadlock by sending
memory by copying if registration fails. It uses the facility introduced
in the fist patch.
Comments are welcome. As always if there will be no complains I'll
commit this to the trunk, so better complain before commit :)