To improve performance on cores with deep pipelines such as ThunderX2,
reimplement gcm(aes) using a 4-way interleave rather than the 2-way
interleave we use currently.
This comes down to a complete rewrite of the GCM part of the combined
GCM/GHASH driver, and instead of interleaving two invocations of AES
with the GHASH handling at the instruction level, the new version
uses a more coarse grained approach where each chunk of 64 bytes is
encrypted first and then ghashed (or ghashed and then decrypted in
the converse case).
The core NEON routine is now able to consume inputs of any size,
and tail blocks of less than 64 bytes are handled using overlapping
loads and stores, and processed by the same 4-way encryption and
hashing routines. This gets rid of most of the branches, and avoids
having to return to the C code to handle the tail block using a
stack buffer.
The table below compares the performance of the old driver and the new
one on various micro-architectures and running in various modes.
| AES-128 | AES-192 | AES-256 |
#bytes | 512 | 1500 | 4k | 512 | 1500 | 4k | 512 | 1500 | 4k |
-------+-----+------+-----+-----+------+-----+-----+------+-----+
TX2 | 35% | 23% | 11% | 34% | 20% | 9% | 38% | 25% | 16% |
EMAG | 11% | 6% | 3% | 12% | 4% | 2% | 11% | 4% | 2% |
A72 | 8% | 5% | -4% | 9% | 4% | -5% | 7% | 4% | -5% |
A53 | 11% | 6% | -1% | 10% | 8% | -1% | 10% | 8% | -2% |
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The GHASH code uses the generic AES key expansion routines, and calls
directly into the scalar table based AES cipher for arm64 from the
fallback path, and since this implementation is known to be non-time
invariant, doing so from a time invariant SIMD cipher is a bit nasty.
So let's switch to the AES library - this makes the code more robust,
and drops the dependency on the generic AES cipher, allowing us to
omit it entirely in the future.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Based on 2 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license version 2 as
published by the free software foundation
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license version 2 as
published by the free software foundation #
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-only
has been chosen to replace the boilerplate/reference in 4122 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Enrico Weigelt <info@metux.net>
Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190604081206.933168790@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Pull crypto update from Herbert Xu:
"API:
- Add support for AEAD in simd
- Add fuzz testing to testmgr
- Add panic_on_fail module parameter to testmgr
- Use per-CPU struct instead multiple variables in scompress
- Change verify API for akcipher
Algorithms:
- Convert x86 AEAD algorithms over to simd
- Forbid 2-key 3DES in FIPS mode
- Add EC-RDSA (GOST 34.10) algorithm
Drivers:
- Set output IV with ctr-aes in crypto4xx
- Set output IV in rockchip
- Fix potential length overflow with hashing in sun4i-ss
- Fix computation error with ctr in vmx
- Add SM4 protected keys support in ccree
- Remove long-broken mxc-scc driver
- Add rfc4106(gcm(aes)) cipher support in cavium/nitrox"
* 'linus' of git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6: (179 commits)
crypto: ccree - use a proper le32 type for le32 val
crypto: ccree - remove set but not used variable 'du_size'
crypto: ccree - Make cc_sec_disable static
crypto: ccree - fix spelling mistake "protedcted" -> "protected"
crypto: caam/qi2 - generate hash keys in-place
crypto: caam/qi2 - fix DMA mapping of stack memory
crypto: caam/qi2 - fix zero-length buffer DMA mapping
crypto: stm32/cryp - update to return iv_out
crypto: stm32/cryp - remove request mutex protection
crypto: stm32/cryp - add weak key check for DES
crypto: atmel - remove set but not used variable 'alg_name'
crypto: picoxcell - Use dev_get_drvdata()
crypto: crypto4xx - get rid of redundant using_sd variable
crypto: crypto4xx - use sync skcipher for fallback
crypto: crypto4xx - fix cfb and ofb "overran dst buffer" issues
crypto: crypto4xx - fix ctr-aes missing output IV
crypto: ecrdsa - select ASN1 and OID_REGISTRY for EC-RDSA
crypto: ux500 - use ccflags-y instead of CFLAGS_<basename>.o
crypto: ccree - handle tee fips error during power management resume
crypto: ccree - add function to handle cryptocell tee fips error
...
As we will exhaust the first 32 bits of AT_HWCAP let's start
exposing AT_HWCAP2 to userspace to give us up to 64 caps.
Whilst it's possible to use the remaining 32 bits of AT_HWCAP, we
prefer to expand into AT_HWCAP2 in order to provide a consistent
view to userspace between ILP32 and LP64. However internal to the
kernel we prefer to continue to use the full space of elf_hwcap.
To reduce complexity and allow for future expansion, we now
represent hwcaps in the kernel as ordinals and use a
KERNEL_HWCAP_ prefix. This allows us to support automatic feature
based module loading for all our hwcaps.
We introduce cpu_set_feature to set hwcaps which complements the
existing cpu_have_feature helper. These helpers allow us to clean
up existing direct uses of elf_hwcap and reduce any future effort
required to move beyond 64 caps.
For convenience we also introduce cpu_{have,set}_named_feature which
makes use of the cpu_feature macro to allow providing a hwcap name
without a {KERNEL_}HWCAP_ prefix.
Signed-off-by: Andrew Murray <andrew.murray@arm.com>
[will: use const_ilog2() and tweak documentation]
Signed-off-by: Will Deacon <will.deacon@arm.com>
Replace all calls to may_use_simd() in the arm64 crypto code with
crypto_simd_usable(), in order to allow testing the no-SIMD code paths.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The arm64 gcm-aes-ce algorithm is failing the extra crypto self-tests
following my patches to test the !may_use_simd() code paths, which
previously were untested. The problem is that in the !may_use_simd()
case, an odd number of AES blocks can be processed within each step of
the skcipher_walk. However, the skcipher_walk is being done with a
"stride" of 2 blocks and is advanced by an even number of blocks after
each step. This causes the encryption to produce the wrong ciphertext
and authentication tag, and causes the decryption to incorrectly fail.
Fix it by only processing an even number of blocks per step.
Fixes: c2b24c36e0 ("crypto: arm64/aes-gcm-ce - fix scatterwalk API violation")
Fixes: 71e52c278c ("crypto: arm64/aes-ce-gcm - operate on two input blocks at a time")
Cc: <stable@vger.kernel.org> # v4.19+
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The arm64 GHASH implementation either uses 8-bit or 64-bit
polynomial multiplication instructions, since the latter are
faster but not mandatory in the architecture.
Since that prevents us from testing both implementations on the
same system, let's expose both implementations to the crypto API,
with the priorities reflecting that the P64 version is the
preferred one if available.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Commit 71e52c278c ("crypto: arm64/aes-ce-gcm - operate on
two input blocks at a time") modified the granularity at which
the AES/GCM code processes its input to allow subsequent changes
to be applied that improve performance by using aggregation to
process multiple input blocks at once.
For this reason, it doubled the algorithm's 'chunksize' property
to 2 x AES_BLOCK_SIZE, but retained the non-SIMD fallback path that
processes a single block at a time. In some cases, this violates the
skcipher scatterwalk API, by calling skcipher_walk_done() with a
non-zero residue value for a chunk that is expected to be handled
in its entirety. This results in a WARN_ON() to be hit by the TLS
self test code, but is likely to break other user cases as well.
Unfortunately, none of the current test cases exercises this exact
code path at the moment.
Fixes: 71e52c278c ("crypto: arm64/aes-ce-gcm - operate on two ...")
Reported-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Enhance the GHASH implementation that uses 64-bit polynomial
multiplication by adding support for 4-way aggregation. This
more than doubles the performance, from 2.4 cycles per byte
to 1.1 cpb on Cortex-A53.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Checking the TIF_NEED_RESCHED flag is disproportionately costly on cores
with fast crypto instructions and comparatively slow memory accesses.
On algorithms such as GHASH, which executes at ~1 cycle per byte on
cores that implement support for 64 bit polynomial multiplication,
there is really no need to check the TIF_NEED_RESCHED particularly
often, and so we can remove the NEON yield check from the assembler
routines.
However, unlike the AEAD or skcipher APIs, the shash/ahash APIs take
arbitrary input lengths, and so there needs to be some sanity check
to ensure that we don't hog the CPU for excessive amounts of time.
So let's simply cap the maximum input size that is processed in one go
to 64 KB.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Squeeze out another 5% of performance by minimizing the number
of invocations of kernel_neon_begin()/kernel_neon_end() on the
common path, which also allows some reloads of the key schedule
to be optimized away.
The resulting code runs at 2.3 cycles per byte on a Cortex-A53.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Implement a faster version of the GHASH transform which amortizes
the reduction modulo the characteristic polynomial across two
input blocks at a time.
On a Cortex-A53, the gcm(aes) performance increases 24%, from
3.0 cycles per byte to 2.4 cpb for large input sizes.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Update the core AES/GCM transform and the associated plumbing to operate
on 2 AES/GHASH blocks at a time. By itself, this is not expected to
result in a noticeable speedup, but it paves the way for reimplementing
the GHASH component using 2-way aggregation.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Calling pmull_gcm_encrypt_block() requires kernel_neon_begin() and
kernel_neon_end() to be used since the routine touches the NEON
register file. Add the missing calls.
Also, since NEON register contents are not preserved outside of
a kernel mode NEON region, pass the key schedule array again.
Fixes: 7c50136a8a ("crypto: arm64/aes-ghash - yield NEON after every ...")
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Many shash algorithms set .cra_flags = CRYPTO_ALG_TYPE_SHASH. But this
is redundant with the C structure type ('struct shash_alg'), and
crypto_register_shash() already sets the type flag automatically,
clearing any type flag that was already there. Apparently the useless
assignment has just been copy+pasted around.
So, remove the useless assignment from all the shash algorithms.
This patch shouldn't change any actual behavior.
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Avoid excessive scheduling delays under a preemptible kernel by
yielding the NEON after every block of input.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Implement a NEON fallback for systems that do support NEON but have
no support for the optional 64x64->128 polynomial multiplication
instruction that is part of the ARMv8 Crypto Extensions. It is based
on the paper "Fast Software Polynomial Multiplication on ARM Processors
Using the NEON Engine" by Danilo Camara, Conrado Gouvea, Julio Lopez and
Ricardo Dahab (https://hal.inria.fr/hal-01506572), but has been reworked
extensively for the AArch64 ISA.
On a low-end core such as the Cortex-A53 found in the Raspberry Pi3, the
NEON based implementation is 4x faster than the table based one, and
is time invariant as well, making it less vulnerable to timing attacks.
When combined with the bit-sliced NEON implementation of AES-CTR, the
AES-GCM performance increases by 2x (from 58 to 29 cycles per byte).
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Currently, the AES-GCM implementation for arm64 systems that support the
ARMv8 Crypto Extensions is based on the generic GCM module, which combines
the AES-CTR implementation using AES instructions with the PMULL based
GHASH driver. This is suboptimal, given the fact that the input data needs
to be loaded twice, once for the encryption and again for the MAC
calculation.
On Cortex-A57 (r1p2) and other recent cores that implement micro-op fusing
for the AES instructions, AES executes at less than 1 cycle per byte, which
means that any cycles wasted on loading the data twice hurt even more.
So implement a new GCM driver that combines the AES and PMULL instructions
at the block level. This improves performance on Cortex-A57 by ~37% (from
3.5 cpb to 2.6 cpb)
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The arm64 kernel will shortly disallow nested kernel mode NEON, so
add a fallback to scalar C code that can be invoked in that case.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This patches modifies the GHASH secure hash implementation to switch to a
faster, polynomial multiplication based reduction instead of one that uses
shifts and rotates.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This fixes a bug in the GHASH algorithm resulting in the calculated hash to be
incorrect if the input is presented in chunks whose size is not a multiple of
16 bytes.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Fixes: fdd2389457 ("arm64/crypto: GHASH secure hash using ARMv8 Crypto Extensions")
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
This is a port to ARMv8 (Crypto Extensions) of the Intel implementation of the
GHASH Secure Hash (used in the Galois/Counter chaining mode). It relies on the
optional PMULL/PMULL2 instruction (polynomial multiply long, what Intel call
carry-less multiply).
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Ard Biesheuvel
Thomas Gleixner
Linus Torvalds
Andrew Murray
Eric Biggers
Herbert Xu