There's no need for the K_table to be made of 64-bit words. For some
reason, the original authors didn't fully reduce the values modulo the
CRC32C polynomial, and so had some 33-bit values in there. They can
all be reduced to 32 bits.
Doing that cuts the table size in half. Since the code depends on both
pclmulq and crc32, SSE 4.1 is obviously present, so we can use pmovzxdq
to fetch it in the correct format.
This adds (measured on Ivy Bridge) 1 cycle per main loop iteration
(CRC of up to 3K bytes), less than 0.2%. The hope is that the reduced
D-cache footprint will make up the loss in other code.
Two other related fixes:
* K_table is read-only, so belongs in .rodata, and
* There's no need for more than 8-byte alignment
Acked-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: George Spelvin <linux@horizon.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The internal key isn't actually in big-endian format so let's switch
to u128 which also happens to allow us to remove a sparse warning.
Based on suggestion by Ard Biesheuvel.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
The GHASH setkey() function uses SSE registers but fails to call
kernel_fpu_begin()/kernel_fpu_end(). Instead of adding these calls, and
then having to deal with the restriction that they cannot be called from
interrupt context, move the setkey() implementation to the C domain.
Note that setkey() does not use any particular SSE features and is not
expected to become a performance bottleneck.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: H. Peter Anvin <hpa@linux.intel.com>
Fixes: 0e1227d356 (crypto: ghash - Add PCLMULQDQ accelerated implementation)
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
There is really no need to page align sha1_transform_avx2. The default
alignment is just fine. This is not the hot code but only the entry
point, after all.
Cc: Chandramouli Narayanan <mouli@linux.intel.com>
Signed-off-by: Mathias Krause <minipli@googlemail.com>
Reviewed-by: H. Peter Anvin <hpa@linux.intel.com>
Reviewed-by: Marek Vasut <marex@denx.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The AVX2 implementation might waste up to a page of stack memory because
of a wrong alignment calculation. This will, in the worst case, increase
the stack usage of sha1_transform_avx2() alone to 5.4 kB -- way to big
for a kernel function. Even worse, it might also allocate *less* bytes
than needed if the stack pointer is already aligned bacause in that case
the 'sub %rbx, %rsp' is effectively moving the stack pointer upwards,
not downwards.
Fix those issues by changing and simplifying the alignment calculation
to use a 32 byte alignment, the alignment really needed.
Cc: Chandramouli Narayanan <mouli@linux.intel.com>
Signed-off-by: Mathias Krause <minipli@googlemail.com>
Reviewed-by: H. Peter Anvin <hpa@linux.intel.com>
Reviewed-by: Marek Vasut <marex@denx.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Commit 7c1da8d0d0 "crypto: sha - SHA1 transform x86_64 AVX2"
accidentally disabled the AVX variant by making the avx_usable() test
not only fail in case the CPU doesn't support AVX or OSXSAVE but also
if it doesn't support AVX2.
Fix that regression by splitting up the AVX/AVX2 test into two
functions. Also test for the BMI1 extension in the avx2_usable() test
as the AVX2 implementation not only makes use of BMI2 but also BMI1
instructions.
Cc: Chandramouli Narayanan <mouli@linux.intel.com>
Signed-off-by: Mathias Krause <minipli@googlemail.com>
Reviewed-by: H. Peter Anvin <hpa@linux.intel.com>
Reviewed-by: Marek Vasut <marex@denx.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This git patch adds x86_64 AVX2 optimization of SHA1
transform to crypto support. The patch has been tested with 3.14.0-rc1
kernel.
On a Haswell desktop, with turbo disabled and all cpus running
at maximum frequency, tcrypt shows AVX2 performance improvement
from 3% for 256 bytes update to 16% for 1024 bytes update over
AVX implementation.
This patch adds sha1_avx2_transform(), the glue, build and
configuration changes needed for AVX2 optimization of
SHA1 transform to crypto support.
sha1-ssse3 is one module which adds the necessary optimization
support (SSSE3/AVX/AVX2) for the low-level SHA1 transform function.
With better optimization support, transform function is overridden
as the case may be. In the case of AVX2, due to performance reasons
across datablock sizes, the AVX or AVX2 transform function is used
at run-time as it suits best. The Makefile change therefore appends
the necessary objects to the linkage. Due to this, the patch merely
appends AVX2 transform to the existing build mix and Kconfig support
and leaves the configuration build support as is.
Signed-off-by: Chandramouli Narayanan <mouli@linux.intel.com>
Reviewed-by: Marek Vasut <marex@denx.de>
Acked-by: H. Peter Anvin <hpa@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
We checked "nbytes < bsize" before so it can't happen here.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Acked-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Acked-by: Johannes Götzfried <johannes.goetzfried@cs.fau.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
We rename aesni-intel_avx.S to aesni-intel_avx-x86_64.S to indicate
that it is only used by x86_64 architecture.
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
It seems commit d764593a "crypto: aesni - AVX and AVX2 version of AESNI-GCM
encode and decode" breaks a build on x86_32 since it's designed only for
x86_64. This patch makes a compilation unit conditional to CONFIG_64BIT and
functions usage to CONFIG_X86_64.
Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
We have added AVX and AVX2 routines that optimize AESNI-GCM encode/decode.
These routines are optimized for encrypt and decrypt of large buffers.
In tests we have seen up to 6% speedup for 1K, 11% speedup for 2K and
18% speedup for 8K buffer over the existing SSE version. These routines
should provide even better speedup for future Intel x86_64 cpus.
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Replace remaining occurences (just as we did in crypto/) under arch/*/crypto/
that make use of memcmp() for comparing keys or authentication tags for
usage with crypto_memneq(). It can simply be used as a drop-in replacement
for the normal memcmp().
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Cc: James Yonan <james@openvpn.net>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The AVX2 implementation also uses BMI2 instructions,
but doesn't test for their availability. The assumption
that AVX2 and BMI2 always go together is false. Some
Haswells have AVX2 but not BMI2.
Signed-off-by: Oliver Neukum <oneukum@suse.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Move all users of ablk_helper under x86/ to the generic version
and delete the x86 specific version.
Acked-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Commit 3d387ef08c (Revert "crypto: blowfish - add AVX2/x86_64 implementation
of blowfish cipher") reverted too much as it removed the 'assembler supports
AVX2' check and therefore disabled remaining AVX2 implementations of Camellia
and Serpent. Patch restores the check and enables these implementations.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Commit a710f761f (crypto: sha256_ssse3 - add sha224 support) attempted to add
MODULE_ALIAS for SHA-224, but it ended up being "sha384", probably because
mix-up with previous commit 340991e30 (crypto: sha512_ssse3 - add sha384
support). Patch corrects module alias to "sha224".
Reported-by: Pierre-Mayeul Badaire <pierre-mayeul.badaire@m4x.org>
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This patch reinstates commits
67822649d739761214ee0b95a7f85731d939625a2d31e518a4
Now that module softdeps are in the kernel we can use that to resolve
the boot issue which cause the revert.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Adjust alignment and replace commas by semicolons in automatically
generated code.
Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Tables used from assembler should be marked __visible to let
the compiler know.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Pull crypto fixes from Herbert Xu:
"This push fixes a memory corruption issue in caam, as well as
reverting the new optimised crct10dif implementation as it breaks boot
on initrd systems.
Hopefully crct10dif will be reinstated once the supporting code is
added so that it doesn't break boot"
* git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6:
Revert "crypto: crct10dif - Wrap crc_t10dif function all to use crypto transform framework"
crypto: caam - Fixed the memory out of bound overwrite issue
This reverts commits
67822649d739761214ee0b95a7f85731d939625a2d31e518a4
Unfortunately this change broke boot on some systems that used an
initrd which does not include the newly created crct10dif modules.
As these modules are required by sd_mod under certain configurations
this is a serious problem.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Pull crypto fix from Herbert Xu:
"This fixes an unaligned crash in XTS mode when using aseni_intel"
* git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6:
crypto: aesni_intel - fix accessing of unaligned memory
This reverts commit cf1521a1a5.
Instruction (vpgatherdd) that this implementation relied on turned out to be
slow performer on real hardware (i5-4570). The previous 8-way twofish/AVX
implementation is therefore faster and this implementation should be removed.
Converting this implementation to use the same method as in twofish/AVX for
table look-ups would give additional ~3% speed up vs twofish/AVX, but would
hardly be worth of the added code and binary size.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This reverts commit 6048801070.
Instruction (vpgatherdd) that this implementation relied on turned out to be
slow performer on real hardware (i5-4570). The previous 4-way blowfish
implementation is therefore faster and this implementation should be removed.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add implementation tuned for more performance on real hardware. Changes are
mostly around the part mixing 128-bit extract and insert instructions and
AES-NI instructions. Also 'vpbroadcastb' instructions have been change to
'vpshufb with zero mask'.
Tests on Intel Core i5-4570:
tcrypt ECB results, old-AVX2 vs new-AVX2:
size 128bit key 256bit key
enc dec enc dec
256 1.00x 1.00x 1.00x 1.00x
1k 1.08x 1.09x 1.05x 1.06x
8k 1.06x 1.06x 1.06x 1.06x
tcrypt ECB results, AVX vs new-AVX2:
size 128bit key 256bit key
enc dec enc dec
256 1.00x 1.00x 1.00x 1.00x
1k 1.51x 1.50x 1.52x 1.50x
8k 1.47x 1.48x 1.48x 1.48x
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The new XTS code for aesni_intel uses input buffers directly as memory operands
for pxor instructions, which causes crash if those buffers are not aligned to
16 bytes.
Patch changes XTS code to handle unaligned memory correctly, by loading memory
with movdqu instead.
Reported-by: Dave Jones <davej@redhat.com>
Tested-by: Dave Jones <davej@redhat.com>
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Pull x86 fixes from Peter Anvin:
- Three EFI-related fixes
- Two early memory initialization fixes
- build fix for older binutils
- fix for an eager FPU performance regression -- currently we don't
allow the use of the FPU at interrupt time *at all* in eager mode,
which is clearly wrong.
* 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86: Allow FPU to be used at interrupt time even with eagerfpu
x86, crc32-pclmul: Fix build with older binutils
x86-64, init: Fix a possible wraparound bug in switchover in head_64.S
x86, range: fix missing merge during add range
x86, efi: initial the local variable of DataSize to zero
efivar: fix oops in efivar_update_sysfs_entries() caused by memory reuse
efivarfs: Never return ENOENT from firmware again
binutils prior to 2.18 (e.g. the ones found on SLE10) don't support
assembling PEXTRD, so a macro based approach like the one for PCLMULQDQ
in the same file should be used.
This requires making the helper macros capable of recognizing 32-bit
general purpose register operands.
[ hpa: tagging for stable as it is a low risk build fix ]
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Link: http://lkml.kernel.org/r/51A6142A02000078000D99D8@nat28.tlf.novell.com
Cc: Alexander Boyko <alexander_boyko@xyratex.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Huang Ying <ying.huang@intel.com>
Cc: <stable@vger.kernel.org> v3.9
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Add sha224 implementation to sha256_ssse3 module.
This also fixes sha256_ssse3 module autoloading issue when 'sha224' is used
before 'sha256'. Previously in such case, just sha256_generic was loaded and
not sha256_ssse3 (since it did not provide sha224). Now if 'sha256' was used
after 'sha224' usage, sha256_ssse3 would remain unloaded.
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add sha384 implementation to sha512_ssse3 module.
This also fixes sha512_ssse3 module autoloading issue when 'sha384' is used
before 'sha512'. Previously in such case, just sha512_generic was loaded and
not sha512_ssse3 (since it did not provide sha384). Now if 'sha512' was used
after 'sha384' usage, sha512_ssse3 would remain unloaded. For example, this
happens with tcrypt testing module since it tests 'sha384' before 'sha512'.
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The _XFER stack element size was set too small, 8 bytes, when it needs to be
16 bytes. As _XFER is the last stack element used by these implementations,
the 16 byte stores with 'movdqa' corrupt the stack where the value of register
%r12 is temporarily stored. As these implementations align the stack pointer
to 16 bytes, this corruption did not happen every time.
Patch corrects this issue.
Reported-by: Julian Wollrath <jwollrath@web.de>
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Tested-by: Julian Wollrath <jwollrath@web.de>
Acked-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Glue code that plugs the PCLMULQDQ accelerated CRC T10 DIF hash into the
crypto framework. The config CRYPTO_CRCT10DIF_PCLMUL should be turned
on to enable the feature. The crc_t10dif crypto library function will
use this faster algorithm when crct10dif_pclmul module is loaded.
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Patch adds AVX2/AES-NI/x86-64 implementation of Camellia cipher, requiring
32 parallel blocks for input (512 bytes). Compared to AVX implementation, this
version is extended to use the 256-bit wide YMM registers. For AES-NI
instructions data is split to two 128-bit registers and merged afterwards.
Even with this additional handling, performance should be higher compared
to the AES-NI/AVX implementation.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Patch adds AVX2/x86-64 implementation of Serpent cipher, requiring 16 parallel
blocks for input (256 bytes). Implementation is based on the AVX implementation
and extends to use the 256-bit wide YMM registers. Since serpent does not use
table look-ups, this implementation should be close to two times faster than
the AVX implementation.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Patch adds AVX2/x86-64 implementation of Twofish cipher, requiring 16 parallel
blocks for input (256 bytes). Table look-ups are performed using vpgatherdd
instruction directly from vector registers and thus should be faster than
earlier implementations. Implementation also uses 256-bit wide YMM registers,
which should give additional speed up compared to the AVX implementation.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Patch adds AVX2/x86-64 implementation of Blowfish cipher, requiring 32 parallel
blocks for input (256 bytes). Table look-ups are performed using vpgatherdd
instruction directly from vector registers and thus should be faster than
earlier implementations.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add more optimized XTS code for aesni_intel in 64-bit mode, for smaller stack
usage and boost for speed.
tcrypt results, with Intel i5-2450M:
256-bit key
enc dec
16B 0.98x 0.99x
64B 0.64x 0.63x
256B 1.29x 1.32x
1024B 1.54x 1.58x
8192B 1.57x 1.60x
512-bit key
enc dec
16B 0.98x 0.99x
64B 0.60x 0.59x
256B 1.24x 1.25x
1024B 1.39x 1.42x
8192B 1.38x 1.42x
I chose not to optimize smaller than block size of 256 bytes, since XTS is
practically always used with data blocks of size 512 bytes. This is why
performance is reduced in tcrypt for 64 byte long blocks.
Cc: Huang Ying <ying.huang@intel.com>
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Add more optimized XTS code for camellia-aesni-avx, for smaller stack usage
and small boost for speed.
tcrypt results, with Intel i5-2450M:
enc dec
16B 1.10x 1.01x
64B 0.82x 0.77x
256B 1.14x 1.10x
1024B 1.17x 1.16x
8192B 1.10x 1.11x
Since XTS is practically always used with data blocks of size 512 bytes or
more, I chose to not make use of camellia-2way for block sized smaller than
256 bytes. This causes slower result in tcrypt for 64 bytes.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Change cast6-avx to use the new XTS code, for smaller stack usage and small
boost to performance.
tcrypt results, with Intel i5-2450M:
enc dec
16B 1.01x 1.01x
64B 1.01x 1.00x
256B 1.09x 1.02x
1024B 1.08x 1.06x
8192B 1.08x 1.07x
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Change twofish-avx to use the new XTS code, for smaller stack usage and small
boost to performance.
tcrypt results, with Intel i5-2450M:
enc dec
16B 1.03x 1.02x
64B 0.91x 0.91x
256B 1.10x 1.09x
1024B 1.12x 1.11x
8192B 1.12x 1.11x
Since XTS is practically always used with data blocks of size 512 bytes or
more, I chose to not make use of twofish-3way for block sized smaller than
128 bytes. This causes slower result in tcrypt for 64 bytes.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This patch adds AVX optimized XTS-mode helper functions/macros and converts
serpent-avx to use the new facilities. Benefits are slightly improved speed
and reduced stack usage as use of temporary IV-array is avoided.
tcrypt results, with Intel i5-2450M:
enc dec
16B 1.00x 1.00x
64B 1.00x 1.00x
256B 1.04x 1.06x
1024B 1.09x 1.09x
8192B 1.10x 1.09x
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Occurs when CONFIG_CRYPTO_CRC32C_INTEL=y and CONFIG_CRYPTO_CRC32C_INTEL=y.
Older versions of bintuils do not support the pclmulqdq instruction. The
PCLMULQDQ gas macro is used instead.
Signed-off-by: Sandy Wu <sandyw@twitter.com>
Cc: stable@vger.kernel.org # 3.8+
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
We added glue code and config options to create crypto
module that uses SSE/AVX/AVX2 optimized SHA512 x86_64 assembly routines.
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Provides SHA512 x86_64 assembly routine optimized with SSE, AVX and
AVX2's RORX instructions. Speedup of 70% or more has been
measured over the generic implementation.
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Provides SHA512 x86_64 assembly routine optimized with SSE and AVX instructions.
Speedup of 60% or more has been measured over the generic implementation.
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
