The previous implementation of DCRYPTO_app_cipher
consumed roughly 16ms to cipher a 16kB buffer
(i.e. performance that is far worse than the
hardware is capable of).
This change speeds up the implementation by about
85%, to the tune of roughly 2.2ms for a 16kB buffer.
The gains originate from various sources: loop
unrolling, data-pipelining, eliminating local
variables (to reduce register pressure), eliminating
support for unaligned input/output data, compiling
hot code with -O (rather the default -Os), and
using the hidden key-ladder, which need only be
setup once per reset.
This change also switches from AES-128 to AES-256.
BRANCH=none
BUG=chrome-os-partner:62260
TEST=make buildall succeeds;
cipher command succeeds;
TCG tests pass
Change-Id: I133741be6d9f1353d6ae732d0e863b4b18cc8c9e
Signed-off-by: nagendra modadugu <ngm@google.com>
Reviewed-on: https://chromium-review.googlesource.com/433359
Commit-Ready: Nagendra Modadugu <ngm@google.com>
Tested-by: Nagendra Modadugu <ngm@google.com>
Reviewed-by: Vadim Bendebury <vbendeb@chromium.org>
On boards based on the g chip cryptographic functions come from
hardware, they should be implemented in chip/g as opposed to a
particular board.
The common modules (like nvmem) should be using some generic API,
which hopefully will be implemented by other chips, or could be
replaced by a purely software implementation where crypto hardware
support is not available.
Crypto API definition is being added in include/ and the g chip
implementation (a wrapper around dcrypto functions) is being added in
chip/g.
test/nvmem_vars.h needed to be edited to avoid conflict with
<string.h>.
BRANCH=none
BUG=chrome-os-partner:62260
TEST=make buildall -j still passes. Booting reef with the new image
works fine too.
Change-Id: Ifef281215f89239966882ecbe3e90c8351b9b91a
Signed-off-by: Vadim Bendebury <vbendeb@chromium.org>
Reviewed-on: https://chromium-review.googlesource.com/431313
Reviewed-by: Randall Spangler <rspangler@chromium.org>
Reviewed-by: Nagendra Modadugu <ngm@google.com>
