Straightforward(ish) port from aesni, without unrolling (block8).
Quick speed test: dd if=/dev/zero of=/dev/md0.eli bs=1m reports ~385 MB/s instead of ~60 MB/s on a Cortex-A72 @ 2GHz (Marvell Armada8k).
Details
Details
- Reviewers
jmg andrew emaste jhb - Group Reviewers
arm64 Contributor Reviews (src) - Commits
- rG4979620ece98: armv8crypto: add AES-XTS support
Tested with opencrypto tests. (D21018 enables them on aarch64)
Diff Detail
Diff Detail
- Repository
- rG FreeBSD src repository
- Lint
Lint Not Applicable - Unit
Tests Not Applicable
Event Timeline
| sys/crypto/armv8/armv8_crypto.c | ||
|---|---|---|
| 188 | Why did you change the argument here ? | |
| sys/crypto/armv8/armv8_crypto.c | ||
|---|---|---|
| 188 | Look at the calls — in the CRD_F_KEY_EXPLICIT case, instead of cryptoini, we have cryptodesc. This is how aesni deals with this. | |
Comment Actions
Rebased on top of the crypto rework that just landed (D23677). Implemented crp->crp_cipher_key handling.
Comment Actions
I'm sorry that this is taking a while to get reviewed. I'd like to help get it in. Would be willing to rebase the patch and re-upload? It doesn't apply for me.
Comment Actions
Looks fine. You should look at unrolling the loop to 3 or 4 rounds. Looking at the A72 optimization guide, it shows that there is a 3 cycle latency, but throughput of 1. Section 4.10 gives example showing three pairs to achieve max perf.
A72 opt guide:
https://developer.arm.com/documentation/uan0016/a
Comment Actions
It would also be good to test with 'cryptocheck -a all -d armv8crypto0 -z' though the NIST KAT are probably sufficient coverage already for XTS.
Comment Actions
Hi,
I was able to rebase this patch with some very minor tweaks, and verify the results using both cryptocheck and cryptotest.py.
Are there any remaining concerns with the patch? I would like to see this committed in the near future.
Comment Actions
I haven't reviewed the actual XTS bits in armv8_crypto_wrap.c, but the rest all looks good to me.