This patch adds crc32 algorithms to shash crypto api. One is wrapper to
gerneric crc32_le function. Second is crc32 pclmulqdq implementation. It
use hardware provided PCLMULQDQ instruction to accelerate the CRC32 disposal.
This instruction present from Intel Westmere and AMD Bulldozer CPUs.
For intel core i5 I got 450MB/s for table implementation and 2100MB/s
for pclmulqdq implementation.
Signed-off-by: Alexander Boyko <alexander_boyko@xyratex.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This patch adds the crc_pcl function that calculates CRC32C checksum using the
PCLMULQDQ instruction on processors that support this feature. This will
provide speedup over using CRC32 instruction only.
The usage of PCLMULQDQ necessitate the invocation of kernel_fpu_begin and
kernel_fpu_end and incur some overhead. So the new crc_pcl function is only
invoked for buffer size of 512 bytes or more. Larger sized
buffers will expect to see greater speedup. This feature is best used coupled
with eager_fpu which reduces the kernel_fpu_begin/end overhead. For
buffer size of 1K the speedup is around 1.6x and for buffer size greater than
4K, the speedup is around 3x compared to original implementation in crc32c-intel
module. Test was performed on Sandy Bridge based platform with constant frequency
set for cpu.
A white paper detailing the algorithm can be found here:
http://download.intel.com/design/intarch/papers/323405.pdf
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This patch renames the crc32c-intel.c file to crc32c-intel_glue.c file
in preparation for linking with the new crc32c-pcl-intel-asm.S file,
which contains optimized crc32c calculation based on PCLMULQDQ
instruction.
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Now that serpent-sse2 glue code has been made generic, it can be split to
separate module.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Move ablk-* functions to separate module to share common code between cipher
implementations.
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Commit ea4d26ae ("raid5: add AVX optimized RAID5 checksumming")
introduced x86/ arch wide defines for AFLAGS and CFLAGS indicating AVX
support in binutils based on the same test we have in x86/crypto/ right
now. To minimize duplication drop our implementation in favour to the
one in x86/.
Signed-off-by: Mathias Krause <minipli@googlemail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Patch adds 3-way parallel x86_64 assembly implementation of twofish as new
module. New assembler functions crypt data in three blocks chunks, improving
cipher performance on out-of-order CPUs.
Patch has been tested with tcrypt and automated filesystem tests.
Summary of the tcrypt benchmarks:
Twofish 3-way-asm vs twofish asm (128bit 8kb block ECB)
encrypt: 1.3x speed
decrypt: 1.3x speed
Twofish 3-way-asm vs twofish asm (128bit 8kb block CBC)
encrypt: 1.07x speed
decrypt: 1.4x speed
Twofish 3-way-asm vs twofish asm (128bit 8kb block CTR)
encrypt: 1.4x speed
Twofish 3-way-asm vs AES asm (128bit 8kb block ECB)
encrypt: 1.0x speed
decrypt: 1.0x speed
Twofish 3-way-asm vs AES asm (128bit 8kb block CBC)
encrypt: 0.84x speed
decrypt: 1.09x speed
Twofish 3-way-asm vs AES asm (128bit 8kb block CTR)
encrypt: 1.15x speed
Full output:
http://koti.mbnet.fi/axh/kernel/crypto/tcrypt-speed-twofish-3way-asm-x86_64.txthttp://koti.mbnet.fi/axh/kernel/crypto/tcrypt-speed-twofish-asm-x86_64.txthttp://koti.mbnet.fi/axh/kernel/crypto/tcrypt-speed-aes-asm-x86_64.txt
Tests were run on:
vendor_id : AuthenticAMD
cpu family : 16
model : 10
model name : AMD Phenom(tm) II X6 1055T Processor
Also userspace test were run on:
vendor_id : GenuineIntel
cpu family : 6
model : 15
model name : Intel(R) Xeon(R) CPU E7330 @ 2.40GHz
stepping : 11
Userspace test results:
Encryption/decryption of twofish 3-way vs x86_64-asm on AMD Phenom II:
encrypt: 1.27x
decrypt: 1.25x
Encryption/decryption of twofish 3-way vs x86_64-asm on Intel Xeon E7330:
encrypt: 1.36x
decrypt: 1.36x
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Patch adds x86_64 assembly implementation of blowfish. Two set of assembler
functions are provided. First set is regular 'one-block at time'
encrypt/decrypt functions. Second is 'four-block at time' functions that
gain performance increase on out-of-order CPUs. Performance of 4-way
functions should be equal to 1-way functions with in-order CPUs.
Summary of the tcrypt benchmarks:
Blowfish assembler vs blowfish C (256bit 8kb block ECB)
encrypt: 2.2x speed
decrypt: 2.3x speed
Blowfish assembler vs blowfish C (256bit 8kb block CBC)
encrypt: 1.12x speed
decrypt: 2.5x speed
Blowfish assembler vs blowfish C (256bit 8kb block CTR)
encrypt: 2.5x speed
Full output:
http://koti.mbnet.fi/axh/kernel/crypto/tcrypt-speed-blowfish-asm-x86_64.txthttp://koti.mbnet.fi/axh/kernel/crypto/tcrypt-speed-blowfish-c-x86_64.txt
Tests were run on:
vendor_id : AuthenticAMD
cpu family : 16
model : 10
model name : AMD Phenom(tm) II X6 1055T Processor
stepping : 0
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This is an assembler implementation of the SHA1 algorithm using the
Supplemental SSE3 (SSSE3) instructions or, when available, the
Advanced Vector Extensions (AVX).
Testing with the tcrypt module shows the raw hash performance is up to
2.3 times faster than the C implementation, using 8k data blocks on a
Core 2 Duo T5500. For the smalest data set (16 byte) it is still 25%
faster.
Since this implementation uses SSE/YMM registers it cannot safely be
used in every situation, e.g. while an IRQ interrupts a kernel thread.
The implementation falls back to the generic SHA1 variant, if using
the SSE/YMM registers is not possible.
With this algorithm I was able to increase the throughput of a single
IPsec link from 344 Mbit/s to 464 Mbit/s on a Core 2 Quad CPU using
the SSSE3 variant -- a speedup of +34.8%.
Saving and restoring SSE/YMM state might make the actual throughput
fluctuate when there are FPU intensive userland applications running.
For example, meassuring the performance using iperf2 directly on the
machine under test gives wobbling numbers because iperf2 uses the FPU
for each packet to check if the reporting interval has expired (in the
above test I got min/max/avg: 402/484/464 MBit/s).
Using this algorithm on a IPsec gateway gives much more reasonable and
stable numbers, albeit not as high as in the directly connected case.
Here is the result from an RFC 2544 test run with a EXFO Packet Blazer
FTB-8510:
frame size sha1-generic sha1-ssse3 delta
64 byte 37.5 MBit/s 37.5 MBit/s 0.0%
128 byte 56.3 MBit/s 62.5 MBit/s +11.0%
256 byte 87.5 MBit/s 100.0 MBit/s +14.3%
512 byte 131.3 MBit/s 150.0 MBit/s +14.2%
1024 byte 162.5 MBit/s 193.8 MBit/s +19.3%
1280 byte 175.0 MBit/s 212.5 MBit/s +21.4%
1420 byte 175.0 MBit/s 218.7 MBit/s +25.0%
1518 byte 150.0 MBit/s 181.2 MBit/s +20.8%
The throughput for the largest frame size is lower than for the
previous size because the IP packets need to be fragmented in this
case to make there way through the IPsec tunnel.
Signed-off-by: Mathias Krause <minipli@googlemail.com>
Cc: Maxim Locktyukhin <maxim.locktyukhin@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Loading fpu without aesni-intel does nothing. Loading aesni-intel
without fpu causes modes like xts to fail. (Unloading
aesni-intel will restore those modes.)
One solution would be to make aesni-intel depend on fpu, but it
seems cleaner to just combine the modules.
This is probably responsible for bugs like:
https://bugzilla.redhat.com/show_bug.cgi?id=589390
Signed-off-by: Andy Lutomirski <luto@mit.edu>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
PCLMULQDQ is used to accelerate the most time-consuming part of GHASH,
carry-less multiplication. More information about PCLMULQDQ can be
found at:
http://software.intel.com/en-us/articles/carry-less-multiplication-and-its-usage-for-computing-the-gcm-mode/
Because PCLMULQDQ changes XMM state, its usage must be enclosed with
kernel_fpu_begin/end, which can be used only in process context, the
acceleration is implemented as crypto_ahash. That is, request in soft
IRQ context will be defered to the cryptd kernel thread.
Signed-off-by: Huang Ying <ying.huang@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Blkcipher touching FPU need to be enclosed by kernel_fpu_begin() and
kernel_fpu_end(). If they are invoked in cipher algorithm
implementation, they will be invoked for each block, so that
performance will be hurt, because they are "slow" operations. This
patch implements "fpu" template, which makes these operations to be
invoked for each request.
Signed-off-by: Huang Ying <ying.huang@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Intel AES-NI is a new set of Single Instruction Multiple Data (SIMD)
instructions that are going to be introduced in the next generation of
Intel processor, as of 2009. These instructions enable fast and secure
data encryption and decryption, using the Advanced Encryption Standard
(AES), defined by FIPS Publication number 197. The architecture
introduces six instructions that offer full hardware support for
AES. Four of them support high performance data encryption and
decryption, and the other two instructions support the AES key
expansion procedure.
The white paper can be downloaded from:
http://softwarecommunity.intel.com/isn/downloads/intelavx/AES-Instructions-Set_WP.pdf
AES may be used in soft_irq context, but MMX/SSE context can not be
touched safely in soft_irq context. So in_interrupt() is checked, if
in IRQ or soft_irq context, the general x86_64 implementation are used
instead.
Signed-off-by: Huang Ying <ying.huang@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
From NHM processor onward, Intel processors can support hardware accelerated
CRC32c algorithm with the new CRC32 instruction in SSE 4.2 instruction set.
The patch detects the availability of the feature, and chooses the most proper
way to calculate CRC32c checksum.
Byte code instructions are used for compiler compatibility.
No MMX / XMM registers is involved in the implementation.
Signed-off-by: Austin Zhang <austin.zhang@intel.com>
Signed-off-by: Kent Liu <kent.liu@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
There is almost no difference between 32 & 64 bit glue code.
Signed-off-by: Sebastian Siewior <sebastian@breakpoint.cc>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This is the x86-64 version of the Salsa20 stream cipher algorithm. The
original assembly code came from
<http://cr.yp.to/snuffle/salsa20/amd64-3/salsa20.s>. It has been
reformatted for clarity.
Signed-off-by: Tan Swee Heng <thesweeheng@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
This patch contains the salsa20-i586 implementation. The original
assembly code came from
<http://cr.yp.to/snuffle/salsa20/x86-pm/salsa20.s>. I have reformatted
it (added indents) so that it matches the other algorithms in
arch/x86/crypto.
Signed-off-by: Tan Swee Heng <thesweeheng@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
32 bit and 64 bit glue code is using (now) the same
piece code. This patch unifies them.
Signed-off-by: Sebastian Siewior <sebastian@breakpoint.cc>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
