Former versions of this (ARM) branch of this driver reworked the hardware-
readable scatter/gather list to operate as a set of 32-bit integers,
rather than a packed structure of smaller sizes, which cannot burst-read
correctly on a little-endian platform.

Integration of caamhash.c revealed subtle ways in which the ordering of
items written to a hardware s/g list could create bugs, such as the
"final" bit being written to an entry that would later be updated with
a size, inadvertently erasing the bit (e.g. such as sg_to_sec4_sg_last()
before sg_to_sec4_sg()).

Since fields must be ORed in to operate correctly using any order of
operations, changed allocations of the combination of extended descriptor
structs + hardware scatterlists to use kzalloc() instead of kmalloc(), so
as to ensure that residue data would not be ORed in with the correct data.

Signed-off-by: Steve Cornelius <steve.cornelius@freescale.com>
Signed-off-by: Terry Lv <r65388@freescale.com>