Currently armv8crypto copies the scheme used in aesni(9), where payload
data and output buffers are allocated on the fly if the crypto buffer is
not virtually contiguous. This scheme is simple but incurs a lot of
overhead: for an encryption request with a separate output buffer we
have to
- allocate a temporary buffer to hold the payload
- copy input data into the buffer
- copy the encrypted payload to the output buffer
- zero the temporary buffer before freeing it
We have a handy crypto buffer cursor abstraction now, so reimplement the
armv8crypto routines using that instead of temporary buffers. This
introduces some extra complexity, but gallatin@ reports a 10% throughput
improvement with a KTLS workload without additional CPU usage. The
driver still allocates an AAD buffer for AES-GCM if necessary.
Reviewed by: jhb
Tested by: gallatin
Sponsored by: Ampere Computing LLC
Submitted by: Klara Inc.
MFC after: 1 month
Differential Revision: https://reviews.freebsd.org/D28950
This is in preparation for using buffer cursors. No functional change
intended.
Reviewed by: jhb
Sponsored by: Ampere Computing LLC
Submitted by: Klara Inc.
MFC after: 1 month
Differential Revision: https://reviews.freebsd.org/D28948
This file inherits some boilerplate and structure from the analogous
file in aesni(4), aesni_wrap.c. Note the derivation and the copyright
holders of that file.
For example, the AES-XTS bits added in 4979620ece were ported from
aesni(4).
Requested by: jmg
Reviewed by: imp, gnn
MFC after: 3 days
Differential Revision: https://reviews.freebsd.org/D29268
This makes it easier to refactor the GCM code to operate on
crypto_buffer_cursors rather than plain contiguous buffers, with the aim
of minimizing the amount of copying and zeroing done today.
No functional change intended.
Reviewed by: jhb
MFC after: 1 week
Sponsored by: Ampere Computing
Submitted by: Klara, Inc.
Differential Revision: https://reviews.freebsd.org/D28500
- We were only hashing up to the first 16 bytes of the AAD.
- When computing the digest during decryption, handle the case where
len == trailer, i.e., len < AES_BLOCK_LEN, properly.
While here:
- trailer is always smaller than AES_BLOCK_LEN, so remove a pair of
unnecessary modulus operations.
- Replace some byte-by-byte loops with memcpy() and memset() calls.
In particular, zero the full block before copying a partial block into
it since we do that elsewhere and it means that the memset() length is
known at compile time.
Reviewed by: jhb
Sponsored by: Ampere Computing
Submitted by: Klara, Inc.
MFC after: 3 days
Differential Revision: https://reviews.freebsd.org/D28501
A straightforward(ish) port from aesni(4). This implementation does not
perform loop unrolling on the input blocks, so this is left as a future
performance improvement.
Submitted by: Greg V <greg AT unrelenting.technology>
Looks good: jhb, jmg
Tested by: mhorne
Differential Revision: https://reviews.freebsd.org/D21017
on the AES-NI code, and modified as needed for use on ARMv8. When loaded
the driver will check the appropriate field in the id_aa64isar0_el1
register to see if AES is supported, and if so the probe function will
signal the driver should attach.
With this I have seen up to 2000Mb/s from the cryptotest test with a single
thread on a ThunderX Pass 2.0.
Reviewed by: imp
Obtained from: ABT Systems Ltd
MFC after: 1 week
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D8297