Archive-Team
/
zfs
2
0
Fork 0
Code Issues Pull Requests Packages Projects Releases 103 Wiki Activity
zfs/module/icp/asm-x86_64/modes/aesni-gcm-x86_64.S

1262 lines
31 KiB
ArmAsm
Raw Normal View History

ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
# Copyright 2013-2016 The OpenSSL Project Authors. All Rights Reserved.
#
# Licensed under the Apache License 2.0 (the "License"). You may not use
# this file except in compliance with the License. You can obtain a copy
# in the file LICENSE in the source distribution or at
# https://www.openssl.org/source/license.html
#
# ====================================================================
# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
# project. The module is, however, dual licensed under OpenSSL and
# CRYPTOGAMS licenses depending on where you obtain it. For further
# details see http://www.openssl.org/~appro/cryptogams/.
# ====================================================================
#
#
# AES-NI-CTR+GHASH stitch.
#
# February 2013
#
# OpenSSL GCM implementation is organized in such way that its
# performance is rather close to the sum of its streamed components,
# in the context parallelized AES-NI CTR and modulo-scheduled
# PCLMULQDQ-enabled GHASH. Unfortunately, as no stitch implementation
# was observed to perform significantly better than the sum of the
# components on contemporary CPUs, the effort was deemed impossible to
# justify. This module is based on combination of Intel submissions,
# [1] and [2], with MOVBE twist suggested by Ilya Albrekht and Max
# Locktyukhin of Intel Corp. who verified that it reduces shuffles
# pressure with notable relative improvement, achieving 1.0 cycle per
# byte processed with 128-bit key on Haswell processor, 0.74 - on
# Broadwell, 0.63 - on Skylake... [Mentioned results are raw profiled
# measurements for favourable packet size, one divisible by 96.
# Applications using the EVP interface will observe a few percent
# worse performance.]
#
# Knights Landing processes 1 byte in 1.25 cycles (measured with EVP).
#
# [1] http://rt.openssl.org/Ticket/Display.html?id=2900&user=guest&pass=guest
# [2] http://www.intel.com/content/dam/www/public/us/en/documents/software-support/enabling-high-performance-gcm.pdf
# Generated once from
# https://github.com/openssl/openssl/blob/5ffc3324/crypto/modes/asm/aesni-gcm-x86_64.pl
# and modified for ICP. Modification are kept at a bare minimum to ease later
# upstream merges.
#if defined(__x86_64__) && defined(HAVE_AVX) && \
ICP: gcm-avx: Support architectures lacking the MOVBE instruction There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup #9749 Closes #10029
2020-03-17 17:24:38 +00:00
defined(HAVE_AES) && defined(HAVE_PCLMULQDQ)
.extern gcm_avx_can_use_movbe
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
.text
ICP: gcm-avx: Support architectures lacking the MOVBE instruction There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup #9749 Closes #10029
2020-03-17 17:24:38 +00:00
#ifdef HAVE_MOVBE
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
.type _aesni_ctr32_ghash_6x,@function
.align 32
_aesni_ctr32_ghash_6x:
Add some missing cfi frame info in aesni-gcm-x86_64.S While preparing #9749 some .cfi_{start,end}proc directives were missed. Add the missing ones. See upstream https://github.com/openssl/openssl/commit/275a048f Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11101
2020-10-30 22:23:18 +00:00
.cfi_startproc
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
vmovdqu 32(%r11),%xmm2
subq $6,%rdx
vpxor %xmm4,%xmm4,%xmm4
vmovdqu 0-128(%rcx),%xmm15
vpaddb %xmm2,%xmm1,%xmm10
vpaddb %xmm2,%xmm10,%xmm11
vpaddb %xmm2,%xmm11,%xmm12
vpaddb %xmm2,%xmm12,%xmm13
vpaddb %xmm2,%xmm13,%xmm14
vpxor %xmm15,%xmm1,%xmm9
vmovdqu %xmm4,16+8(%rsp)
jmp .Loop6x
.align 32
.Loop6x:
addl $100663296,%ebx
jc .Lhandle_ctr32
vmovdqu 0-32(%r9),%xmm3
vpaddb %xmm2,%xmm14,%xmm1
vpxor %xmm15,%xmm10,%xmm10
vpxor %xmm15,%xmm11,%xmm11
.Lresume_ctr32:
vmovdqu %xmm1,(%r8)
vpclmulqdq $0x10,%xmm3,%xmm7,%xmm5
vpxor %xmm15,%xmm12,%xmm12
vmovups 16-128(%rcx),%xmm2
vpclmulqdq $0x01,%xmm3,%xmm7,%xmm6
xorq %r12,%r12
cmpq %r14,%r15
vaesenc %xmm2,%xmm9,%xmm9
vmovdqu 48+8(%rsp),%xmm0
vpxor %xmm15,%xmm13,%xmm13
vpclmulqdq $0x00,%xmm3,%xmm7,%xmm1
vaesenc %xmm2,%xmm10,%xmm10
vpxor %xmm15,%xmm14,%xmm14
setnc %r12b
vpclmulqdq $0x11,%xmm3,%xmm7,%xmm7
vaesenc %xmm2,%xmm11,%xmm11
vmovdqu 16-32(%r9),%xmm3
negq %r12
vaesenc %xmm2,%xmm12,%xmm12
vpxor %xmm5,%xmm6,%xmm6
vpclmulqdq $0x00,%xmm3,%xmm0,%xmm5
vpxor %xmm4,%xmm8,%xmm8
vaesenc %xmm2,%xmm13,%xmm13
vpxor %xmm5,%xmm1,%xmm4
andq $0x60,%r12
vmovups 32-128(%rcx),%xmm15
vpclmulqdq $0x10,%xmm3,%xmm0,%xmm1
vaesenc %xmm2,%xmm14,%xmm14
vpclmulqdq $0x01,%xmm3,%xmm0,%xmm2
leaq (%r14,%r12,1),%r14
vaesenc %xmm15,%xmm9,%xmm9
vpxor 16+8(%rsp),%xmm8,%xmm8
vpclmulqdq $0x11,%xmm3,%xmm0,%xmm3
vmovdqu 64+8(%rsp),%xmm0
vaesenc %xmm15,%xmm10,%xmm10
movbeq 88(%r14),%r13
vaesenc %xmm15,%xmm11,%xmm11
movbeq 80(%r14),%r12
vaesenc %xmm15,%xmm12,%xmm12
movq %r13,32+8(%rsp)
vaesenc %xmm15,%xmm13,%xmm13
movq %r12,40+8(%rsp)
vmovdqu 48-32(%r9),%xmm5
vaesenc %xmm15,%xmm14,%xmm14
vmovups 48-128(%rcx),%xmm15
vpxor %xmm1,%xmm6,%xmm6
vpclmulqdq $0x00,%xmm5,%xmm0,%xmm1
vaesenc %xmm15,%xmm9,%xmm9
vpxor %xmm2,%xmm6,%xmm6
vpclmulqdq $0x10,%xmm5,%xmm0,%xmm2
vaesenc %xmm15,%xmm10,%xmm10
vpxor %xmm3,%xmm7,%xmm7
vpclmulqdq $0x01,%xmm5,%xmm0,%xmm3
vaesenc %xmm15,%xmm11,%xmm11
vpclmulqdq $0x11,%xmm5,%xmm0,%xmm5
vmovdqu 80+8(%rsp),%xmm0
vaesenc %xmm15,%xmm12,%xmm12
vaesenc %xmm15,%xmm13,%xmm13
vpxor %xmm1,%xmm4,%xmm4
vmovdqu 64-32(%r9),%xmm1
vaesenc %xmm15,%xmm14,%xmm14
vmovups 64-128(%rcx),%xmm15
vpxor %xmm2,%xmm6,%xmm6
vpclmulqdq $0x00,%xmm1,%xmm0,%xmm2
vaesenc %xmm15,%xmm9,%xmm9
vpxor %xmm3,%xmm6,%xmm6
vpclmulqdq $0x10,%xmm1,%xmm0,%xmm3
vaesenc %xmm15,%xmm10,%xmm10
movbeq 72(%r14),%r13
vpxor %xmm5,%xmm7,%xmm7
vpclmulqdq $0x01,%xmm1,%xmm0,%xmm5
vaesenc %xmm15,%xmm11,%xmm11
movbeq 64(%r14),%r12
vpclmulqdq $0x11,%xmm1,%xmm0,%xmm1
vmovdqu 96+8(%rsp),%xmm0
vaesenc %xmm15,%xmm12,%xmm12
movq %r13,48+8(%rsp)
vaesenc %xmm15,%xmm13,%xmm13
movq %r12,56+8(%rsp)
vpxor %xmm2,%xmm4,%xmm4
vmovdqu 96-32(%r9),%xmm2
vaesenc %xmm15,%xmm14,%xmm14
vmovups 80-128(%rcx),%xmm15
vpxor %xmm3,%xmm6,%xmm6
vpclmulqdq $0x00,%xmm2,%xmm0,%xmm3
vaesenc %xmm15,%xmm9,%xmm9
vpxor %xmm5,%xmm6,%xmm6
vpclmulqdq $0x10,%xmm2,%xmm0,%xmm5
vaesenc %xmm15,%xmm10,%xmm10
movbeq 56(%r14),%r13
vpxor %xmm1,%xmm7,%xmm7
vpclmulqdq $0x01,%xmm2,%xmm0,%xmm1
vpxor 112+8(%rsp),%xmm8,%xmm8
vaesenc %xmm15,%xmm11,%xmm11
movbeq 48(%r14),%r12
vpclmulqdq $0x11,%xmm2,%xmm0,%xmm2
vaesenc %xmm15,%xmm12,%xmm12
movq %r13,64+8(%rsp)
vaesenc %xmm15,%xmm13,%xmm13
movq %r12,72+8(%rsp)
vpxor %xmm3,%xmm4,%xmm4
vmovdqu 112-32(%r9),%xmm3
vaesenc %xmm15,%xmm14,%xmm14
vmovups 96-128(%rcx),%xmm15
vpxor %xmm5,%xmm6,%xmm6
vpclmulqdq $0x10,%xmm3,%xmm8,%xmm5
vaesenc %xmm15,%xmm9,%xmm9
vpxor %xmm1,%xmm6,%xmm6
vpclmulqdq $0x01,%xmm3,%xmm8,%xmm1
vaesenc %xmm15,%xmm10,%xmm10
movbeq 40(%r14),%r13
vpxor %xmm2,%xmm7,%xmm7
vpclmulqdq $0x00,%xmm3,%xmm8,%xmm2
vaesenc %xmm15,%xmm11,%xmm11
movbeq 32(%r14),%r12
vpclmulqdq $0x11,%xmm3,%xmm8,%xmm8
vaesenc %xmm15,%xmm12,%xmm12
movq %r13,80+8(%rsp)
vaesenc %xmm15,%xmm13,%xmm13
movq %r12,88+8(%rsp)
vpxor %xmm5,%xmm6,%xmm6
vaesenc %xmm15,%xmm14,%xmm14
vpxor %xmm1,%xmm6,%xmm6
vmovups 112-128(%rcx),%xmm15
vpslldq $8,%xmm6,%xmm5
vpxor %xmm2,%xmm4,%xmm4
vmovdqu 16(%r11),%xmm3
vaesenc %xmm15,%xmm9,%xmm9
vpxor %xmm8,%xmm7,%xmm7
vaesenc %xmm15,%xmm10,%xmm10
vpxor %xmm5,%xmm4,%xmm4
movbeq 24(%r14),%r13
vaesenc %xmm15,%xmm11,%xmm11
movbeq 16(%r14),%r12
vpalignr $8,%xmm4,%xmm4,%xmm0
vpclmulqdq $0x10,%xmm3,%xmm4,%xmm4
movq %r13,96+8(%rsp)
vaesenc %xmm15,%xmm12,%xmm12
movq %r12,104+8(%rsp)
vaesenc %xmm15,%xmm13,%xmm13
vmovups 128-128(%rcx),%xmm1
vaesenc %xmm15,%xmm14,%xmm14
vaesenc %xmm1,%xmm9,%xmm9
vmovups 144-128(%rcx),%xmm15
vaesenc %xmm1,%xmm10,%xmm10
vpsrldq $8,%xmm6,%xmm6
vaesenc %xmm1,%xmm11,%xmm11
vpxor %xmm6,%xmm7,%xmm7
vaesenc %xmm1,%xmm12,%xmm12
vpxor %xmm0,%xmm4,%xmm4
movbeq 8(%r14),%r13
vaesenc %xmm1,%xmm13,%xmm13
movbeq 0(%r14),%r12
vaesenc %xmm1,%xmm14,%xmm14
vmovups 160-128(%rcx),%xmm1
cmpl $12,%ebp // ICP uses 10,12,14 not 9,11,13 for rounds.
jb .Lenc_tail
vaesenc %xmm15,%xmm9,%xmm9
vaesenc %xmm15,%xmm10,%xmm10
vaesenc %xmm15,%xmm11,%xmm11
vaesenc %xmm15,%xmm12,%xmm12
vaesenc %xmm15,%xmm13,%xmm13
vaesenc %xmm15,%xmm14,%xmm14
vaesenc %xmm1,%xmm9,%xmm9
vaesenc %xmm1,%xmm10,%xmm10
vaesenc %xmm1,%xmm11,%xmm11
vaesenc %xmm1,%xmm12,%xmm12
vaesenc %xmm1,%xmm13,%xmm13
vmovups 176-128(%rcx),%xmm15
vaesenc %xmm1,%xmm14,%xmm14
vmovups 192-128(%rcx),%xmm1
cmpl $14,%ebp // ICP does not zero key schedule.
jb .Lenc_tail
vaesenc %xmm15,%xmm9,%xmm9
vaesenc %xmm15,%xmm10,%xmm10
vaesenc %xmm15,%xmm11,%xmm11
vaesenc %xmm15,%xmm12,%xmm12
vaesenc %xmm15,%xmm13,%xmm13
vaesenc %xmm15,%xmm14,%xmm14
vaesenc %xmm1,%xmm9,%xmm9
vaesenc %xmm1,%xmm10,%xmm10
vaesenc %xmm1,%xmm11,%xmm11
vaesenc %xmm1,%xmm12,%xmm12
vaesenc %xmm1,%xmm13,%xmm13
vmovups 208-128(%rcx),%xmm15
vaesenc %xmm1,%xmm14,%xmm14
vmovups 224-128(%rcx),%xmm1
jmp .Lenc_tail
.align 32
.Lhandle_ctr32:
vmovdqu (%r11),%xmm0
vpshufb %xmm0,%xmm1,%xmm6
vmovdqu 48(%r11),%xmm5
vpaddd 64(%r11),%xmm6,%xmm10
vpaddd %xmm5,%xmm6,%xmm11
vmovdqu 0-32(%r9),%xmm3
vpaddd %xmm5,%xmm10,%xmm12
vpshufb %xmm0,%xmm10,%xmm10
vpaddd %xmm5,%xmm11,%xmm13
vpshufb %xmm0,%xmm11,%xmm11
vpxor %xmm15,%xmm10,%xmm10
vpaddd %xmm5,%xmm12,%xmm14
vpshufb %xmm0,%xmm12,%xmm12
vpxor %xmm15,%xmm11,%xmm11
vpaddd %xmm5,%xmm13,%xmm1
vpshufb %xmm0,%xmm13,%xmm13
vpshufb %xmm0,%xmm14,%xmm14
vpshufb %xmm0,%xmm1,%xmm1
jmp .Lresume_ctr32
.align 32
.Lenc_tail:
vaesenc %xmm15,%xmm9,%xmm9
vmovdqu %xmm7,16+8(%rsp)
vpalignr $8,%xmm4,%xmm4,%xmm8
vaesenc %xmm15,%xmm10,%xmm10
vpclmulqdq $0x10,%xmm3,%xmm4,%xmm4
vpxor 0(%rdi),%xmm1,%xmm2
vaesenc %xmm15,%xmm11,%xmm11
vpxor 16(%rdi),%xmm1,%xmm0
vaesenc %xmm15,%xmm12,%xmm12
vpxor 32(%rdi),%xmm1,%xmm5
vaesenc %xmm15,%xmm13,%xmm13
vpxor 48(%rdi),%xmm1,%xmm6
vaesenc %xmm15,%xmm14,%xmm14
vpxor 64(%rdi),%xmm1,%xmm7
vpxor 80(%rdi),%xmm1,%xmm3
vmovdqu (%r8),%xmm1
vaesenclast %xmm2,%xmm9,%xmm9
vmovdqu 32(%r11),%xmm2
vaesenclast %xmm0,%xmm10,%xmm10
vpaddb %xmm2,%xmm1,%xmm0
movq %r13,112+8(%rsp)
leaq 96(%rdi),%rdi
vaesenclast %xmm5,%xmm11,%xmm11
vpaddb %xmm2,%xmm0,%xmm5
movq %r12,120+8(%rsp)
leaq 96(%rsi),%rsi
vmovdqu 0-128(%rcx),%xmm15
vaesenclast %xmm6,%xmm12,%xmm12
vpaddb %xmm2,%xmm5,%xmm6
vaesenclast %xmm7,%xmm13,%xmm13
vpaddb %xmm2,%xmm6,%xmm7
vaesenclast %xmm3,%xmm14,%xmm14
vpaddb %xmm2,%xmm7,%xmm3
addq $0x60,%r10
subq $0x6,%rdx
jc .L6x_done
vmovups %xmm9,-96(%rsi)
vpxor %xmm15,%xmm1,%xmm9
vmovups %xmm10,-80(%rsi)
vmovdqa %xmm0,%xmm10
vmovups %xmm11,-64(%rsi)
vmovdqa %xmm5,%xmm11
vmovups %xmm12,-48(%rsi)
vmovdqa %xmm6,%xmm12
vmovups %xmm13,-32(%rsi)
vmovdqa %xmm7,%xmm13
vmovups %xmm14,-16(%rsi)
vmovdqa %xmm3,%xmm14
vmovdqu 32+8(%rsp),%xmm7
jmp .Loop6x
.L6x_done:
vpxor 16+8(%rsp),%xmm8,%xmm8
vpxor %xmm4,%xmm8,%xmm8
.byte 0xf3,0xc3
Add some missing cfi frame info in aesni-gcm-x86_64.S While preparing #9749 some .cfi_{start,end}proc directives were missed. Add the missing ones. See upstream https://github.com/openssl/openssl/commit/275a048f Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11101
2020-10-30 22:23:18 +00:00
.cfi_endproc
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
.size _aesni_ctr32_ghash_6x,.-_aesni_ctr32_ghash_6x
ICP: gcm-avx: Support architectures lacking the MOVBE instruction There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup #9749 Closes #10029
2020-03-17 17:24:38 +00:00
#endif /* ifdef HAVE_MOVBE */
.type _aesni_ctr32_ghash_no_movbe_6x,@function
.align 32
_aesni_ctr32_ghash_no_movbe_6x:
Add some missing cfi frame info in aesni-gcm-x86_64.S While preparing #9749 some .cfi_{start,end}proc directives were missed. Add the missing ones. See upstream https://github.com/openssl/openssl/commit/275a048f Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11101
2020-10-30 22:23:18 +00:00
.cfi_startproc
ICP: gcm-avx: Support architectures lacking the MOVBE instruction There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup #9749 Closes #10029
2020-03-17 17:24:38 +00:00
vmovdqu 32(%r11),%xmm2
subq $6,%rdx
vpxor %xmm4,%xmm4,%xmm4
vmovdqu 0-128(%rcx),%xmm15
vpaddb %xmm2,%xmm1,%xmm10
vpaddb %xmm2,%xmm10,%xmm11
vpaddb %xmm2,%xmm11,%xmm12
vpaddb %xmm2,%xmm12,%xmm13
vpaddb %xmm2,%xmm13,%xmm14
vpxor %xmm15,%xmm1,%xmm9
vmovdqu %xmm4,16+8(%rsp)
jmp .Loop6x_nmb
.align 32
.Loop6x_nmb:
addl $100663296,%ebx
jc .Lhandle_ctr32_nmb
vmovdqu 0-32(%r9),%xmm3
vpaddb %xmm2,%xmm14,%xmm1
vpxor %xmm15,%xmm10,%xmm10
vpxor %xmm15,%xmm11,%xmm11
.Lresume_ctr32_nmb:
vmovdqu %xmm1,(%r8)
vpclmulqdq $0x10,%xmm3,%xmm7,%xmm5
vpxor %xmm15,%xmm12,%xmm12
vmovups 16-128(%rcx),%xmm2
vpclmulqdq $0x01,%xmm3,%xmm7,%xmm6
xorq %r12,%r12
cmpq %r14,%r15
vaesenc %xmm2,%xmm9,%xmm9
vmovdqu 48+8(%rsp),%xmm0
vpxor %xmm15,%xmm13,%xmm13
vpclmulqdq $0x00,%xmm3,%xmm7,%xmm1
vaesenc %xmm2,%xmm10,%xmm10
vpxor %xmm15,%xmm14,%xmm14
setnc %r12b
vpclmulqdq $0x11,%xmm3,%xmm7,%xmm7
vaesenc %xmm2,%xmm11,%xmm11
vmovdqu 16-32(%r9),%xmm3
negq %r12
vaesenc %xmm2,%xmm12,%xmm12
vpxor %xmm5,%xmm6,%xmm6
vpclmulqdq $0x00,%xmm3,%xmm0,%xmm5
vpxor %xmm4,%xmm8,%xmm8
vaesenc %xmm2,%xmm13,%xmm13
vpxor %xmm5,%xmm1,%xmm4
andq $0x60,%r12
vmovups 32-128(%rcx),%xmm15
vpclmulqdq $0x10,%xmm3,%xmm0,%xmm1
vaesenc %xmm2,%xmm14,%xmm14
vpclmulqdq $0x01,%xmm3,%xmm0,%xmm2
leaq (%r14,%r12,1),%r14
vaesenc %xmm15,%xmm9,%xmm9
vpxor 16+8(%rsp),%xmm8,%xmm8
vpclmulqdq $0x11,%xmm3,%xmm0,%xmm3
vmovdqu 64+8(%rsp),%xmm0
vaesenc %xmm15,%xmm10,%xmm10
movq 88(%r14),%r13
bswapq %r13
vaesenc %xmm15,%xmm11,%xmm11
movq 80(%r14),%r12
bswapq %r12
vaesenc %xmm15,%xmm12,%xmm12
movq %r13,32+8(%rsp)
vaesenc %xmm15,%xmm13,%xmm13
movq %r12,40+8(%rsp)
vmovdqu 48-32(%r9),%xmm5
vaesenc %xmm15,%xmm14,%xmm14
vmovups 48-128(%rcx),%xmm15
vpxor %xmm1,%xmm6,%xmm6
vpclmulqdq $0x00,%xmm5,%xmm0,%xmm1
vaesenc %xmm15,%xmm9,%xmm9
vpxor %xmm2,%xmm6,%xmm6
vpclmulqdq $0x10,%xmm5,%xmm0,%xmm2
vaesenc %xmm15,%xmm10,%xmm10
vpxor %xmm3,%xmm7,%xmm7
vpclmulqdq $0x01,%xmm5,%xmm0,%xmm3
vaesenc %xmm15,%xmm11,%xmm11
vpclmulqdq $0x11,%xmm5,%xmm0,%xmm5
vmovdqu 80+8(%rsp),%xmm0
vaesenc %xmm15,%xmm12,%xmm12
vaesenc %xmm15,%xmm13,%xmm13
vpxor %xmm1,%xmm4,%xmm4
vmovdqu 64-32(%r9),%xmm1
vaesenc %xmm15,%xmm14,%xmm14
vmovups 64-128(%rcx),%xmm15
vpxor %xmm2,%xmm6,%xmm6
vpclmulqdq $0x00,%xmm1,%xmm0,%xmm2
vaesenc %xmm15,%xmm9,%xmm9
vpxor %xmm3,%xmm6,%xmm6
vpclmulqdq $0x10,%xmm1,%xmm0,%xmm3
vaesenc %xmm15,%xmm10,%xmm10
movq 72(%r14),%r13
bswapq %r13
vpxor %xmm5,%xmm7,%xmm7
vpclmulqdq $0x01,%xmm1,%xmm0,%xmm5
vaesenc %xmm15,%xmm11,%xmm11
movq 64(%r14),%r12
bswapq %r12
vpclmulqdq $0x11,%xmm1,%xmm0,%xmm1
vmovdqu 96+8(%rsp),%xmm0
vaesenc %xmm15,%xmm12,%xmm12
movq %r13,48+8(%rsp)
vaesenc %xmm15,%xmm13,%xmm13
movq %r12,56+8(%rsp)
vpxor %xmm2,%xmm4,%xmm4
vmovdqu 96-32(%r9),%xmm2
vaesenc %xmm15,%xmm14,%xmm14
vmovups 80-128(%rcx),%xmm15
vpxor %xmm3,%xmm6,%xmm6
vpclmulqdq $0x00,%xmm2,%xmm0,%xmm3
vaesenc %xmm15,%xmm9,%xmm9
vpxor %xmm5,%xmm6,%xmm6
vpclmulqdq $0x10,%xmm2,%xmm0,%xmm5
vaesenc %xmm15,%xmm10,%xmm10
movq 56(%r14),%r13
bswapq %r13
vpxor %xmm1,%xmm7,%xmm7
vpclmulqdq $0x01,%xmm2,%xmm0,%xmm1
vpxor 112+8(%rsp),%xmm8,%xmm8
vaesenc %xmm15,%xmm11,%xmm11
movq 48(%r14),%r12
bswapq %r12
vpclmulqdq $0x11,%xmm2,%xmm0,%xmm2
vaesenc %xmm15,%xmm12,%xmm12
movq %r13,64+8(%rsp)
vaesenc %xmm15,%xmm13,%xmm13
movq %r12,72+8(%rsp)
vpxor %xmm3,%xmm4,%xmm4
vmovdqu 112-32(%r9),%xmm3
vaesenc %xmm15,%xmm14,%xmm14
vmovups 96-128(%rcx),%xmm15
vpxor %xmm5,%xmm6,%xmm6
vpclmulqdq $0x10,%xmm3,%xmm8,%xmm5
vaesenc %xmm15,%xmm9,%xmm9
vpxor %xmm1,%xmm6,%xmm6
vpclmulqdq $0x01,%xmm3,%xmm8,%xmm1
vaesenc %xmm15,%xmm10,%xmm10
movq 40(%r14),%r13
bswapq %r13
vpxor %xmm2,%xmm7,%xmm7
vpclmulqdq $0x00,%xmm3,%xmm8,%xmm2
vaesenc %xmm15,%xmm11,%xmm11
movq 32(%r14),%r12
bswapq %r12
vpclmulqdq $0x11,%xmm3,%xmm8,%xmm8
vaesenc %xmm15,%xmm12,%xmm12
movq %r13,80+8(%rsp)
vaesenc %xmm15,%xmm13,%xmm13
movq %r12,88+8(%rsp)
vpxor %xmm5,%xmm6,%xmm6
vaesenc %xmm15,%xmm14,%xmm14
vpxor %xmm1,%xmm6,%xmm6
vmovups 112-128(%rcx),%xmm15
vpslldq $8,%xmm6,%xmm5
vpxor %xmm2,%xmm4,%xmm4
vmovdqu 16(%r11),%xmm3
vaesenc %xmm15,%xmm9,%xmm9
vpxor %xmm8,%xmm7,%xmm7
vaesenc %xmm15,%xmm10,%xmm10
vpxor %xmm5,%xmm4,%xmm4
movq 24(%r14),%r13
bswapq %r13
vaesenc %xmm15,%xmm11,%xmm11
movq 16(%r14),%r12
bswapq %r12
vpalignr $8,%xmm4,%xmm4,%xmm0
vpclmulqdq $0x10,%xmm3,%xmm4,%xmm4
movq %r13,96+8(%rsp)
vaesenc %xmm15,%xmm12,%xmm12
movq %r12,104+8(%rsp)
vaesenc %xmm15,%xmm13,%xmm13
vmovups 128-128(%rcx),%xmm1
vaesenc %xmm15,%xmm14,%xmm14
vaesenc %xmm1,%xmm9,%xmm9
vmovups 144-128(%rcx),%xmm15
vaesenc %xmm1,%xmm10,%xmm10
vpsrldq $8,%xmm6,%xmm6
vaesenc %xmm1,%xmm11,%xmm11
vpxor %xmm6,%xmm7,%xmm7
vaesenc %xmm1,%xmm12,%xmm12
vpxor %xmm0,%xmm4,%xmm4
movq 8(%r14),%r13
bswapq %r13
vaesenc %xmm1,%xmm13,%xmm13
movq 0(%r14),%r12
bswapq %r12
vaesenc %xmm1,%xmm14,%xmm14
vmovups 160-128(%rcx),%xmm1
cmpl $12,%ebp // ICP uses 10,12,14 not 9,11,13 for rounds.
jb .Lenc_tail_nmb
vaesenc %xmm15,%xmm9,%xmm9
vaesenc %xmm15,%xmm10,%xmm10
vaesenc %xmm15,%xmm11,%xmm11
vaesenc %xmm15,%xmm12,%xmm12
vaesenc %xmm15,%xmm13,%xmm13
vaesenc %xmm15,%xmm14,%xmm14
vaesenc %xmm1,%xmm9,%xmm9
vaesenc %xmm1,%xmm10,%xmm10
vaesenc %xmm1,%xmm11,%xmm11
vaesenc %xmm1,%xmm12,%xmm12
vaesenc %xmm1,%xmm13,%xmm13
vmovups 176-128(%rcx),%xmm15
vaesenc %xmm1,%xmm14,%xmm14
vmovups 192-128(%rcx),%xmm1
cmpl $14,%ebp // ICP does not zero key schedule.
jb .Lenc_tail_nmb
vaesenc %xmm15,%xmm9,%xmm9
vaesenc %xmm15,%xmm10,%xmm10
vaesenc %xmm15,%xmm11,%xmm11
vaesenc %xmm15,%xmm12,%xmm12
vaesenc %xmm15,%xmm13,%xmm13
vaesenc %xmm15,%xmm14,%xmm14
vaesenc %xmm1,%xmm9,%xmm9
vaesenc %xmm1,%xmm10,%xmm10
vaesenc %xmm1,%xmm11,%xmm11
vaesenc %xmm1,%xmm12,%xmm12
vaesenc %xmm1,%xmm13,%xmm13
vmovups 208-128(%rcx),%xmm15
vaesenc %xmm1,%xmm14,%xmm14
vmovups 224-128(%rcx),%xmm1
jmp .Lenc_tail_nmb
.align 32
.Lhandle_ctr32_nmb:
vmovdqu (%r11),%xmm0
vpshufb %xmm0,%xmm1,%xmm6
vmovdqu 48(%r11),%xmm5
vpaddd 64(%r11),%xmm6,%xmm10
vpaddd %xmm5,%xmm6,%xmm11
vmovdqu 0-32(%r9),%xmm3
vpaddd %xmm5,%xmm10,%xmm12
vpshufb %xmm0,%xmm10,%xmm10
vpaddd %xmm5,%xmm11,%xmm13
vpshufb %xmm0,%xmm11,%xmm11
vpxor %xmm15,%xmm10,%xmm10
vpaddd %xmm5,%xmm12,%xmm14
vpshufb %xmm0,%xmm12,%xmm12
vpxor %xmm15,%xmm11,%xmm11
vpaddd %xmm5,%xmm13,%xmm1
vpshufb %xmm0,%xmm13,%xmm13
vpshufb %xmm0,%xmm14,%xmm14
vpshufb %xmm0,%xmm1,%xmm1
jmp .Lresume_ctr32_nmb
.align 32
.Lenc_tail_nmb:
vaesenc %xmm15,%xmm9,%xmm9
vmovdqu %xmm7,16+8(%rsp)
vpalignr $8,%xmm4,%xmm4,%xmm8
vaesenc %xmm15,%xmm10,%xmm10
vpclmulqdq $0x10,%xmm3,%xmm4,%xmm4
vpxor 0(%rdi),%xmm1,%xmm2
vaesenc %xmm15,%xmm11,%xmm11
vpxor 16(%rdi),%xmm1,%xmm0
vaesenc %xmm15,%xmm12,%xmm12
vpxor 32(%rdi),%xmm1,%xmm5
vaesenc %xmm15,%xmm13,%xmm13
vpxor 48(%rdi),%xmm1,%xmm6
vaesenc %xmm15,%xmm14,%xmm14
vpxor 64(%rdi),%xmm1,%xmm7
vpxor 80(%rdi),%xmm1,%xmm3
vmovdqu (%r8),%xmm1
vaesenclast %xmm2,%xmm9,%xmm9
vmovdqu 32(%r11),%xmm2
vaesenclast %xmm0,%xmm10,%xmm10
vpaddb %xmm2,%xmm1,%xmm0
movq %r13,112+8(%rsp)
leaq 96(%rdi),%rdi
vaesenclast %xmm5,%xmm11,%xmm11
vpaddb %xmm2,%xmm0,%xmm5
movq %r12,120+8(%rsp)
leaq 96(%rsi),%rsi
vmovdqu 0-128(%rcx),%xmm15
vaesenclast %xmm6,%xmm12,%xmm12
vpaddb %xmm2,%xmm5,%xmm6
vaesenclast %xmm7,%xmm13,%xmm13
vpaddb %xmm2,%xmm6,%xmm7
vaesenclast %xmm3,%xmm14,%xmm14
vpaddb %xmm2,%xmm7,%xmm3
addq $0x60,%r10
subq $0x6,%rdx
jc .L6x_done_nmb
vmovups %xmm9,-96(%rsi)
vpxor %xmm15,%xmm1,%xmm9
vmovups %xmm10,-80(%rsi)
vmovdqa %xmm0,%xmm10
vmovups %xmm11,-64(%rsi)
vmovdqa %xmm5,%xmm11
vmovups %xmm12,-48(%rsi)
vmovdqa %xmm6,%xmm12
vmovups %xmm13,-32(%rsi)
vmovdqa %xmm7,%xmm13
vmovups %xmm14,-16(%rsi)
vmovdqa %xmm3,%xmm14
vmovdqu 32+8(%rsp),%xmm7
jmp .Loop6x_nmb
.L6x_done_nmb:
vpxor 16+8(%rsp),%xmm8,%xmm8
vpxor %xmm4,%xmm8,%xmm8
.byte 0xf3,0xc3
Add some missing cfi frame info in aesni-gcm-x86_64.S While preparing #9749 some .cfi_{start,end}proc directives were missed. Add the missing ones. See upstream https://github.com/openssl/openssl/commit/275a048f Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11101
2020-10-30 22:23:18 +00:00
.cfi_endproc
ICP: gcm-avx: Support architectures lacking the MOVBE instruction There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup #9749 Closes #10029
2020-03-17 17:24:38 +00:00
.size _aesni_ctr32_ghash_no_movbe_6x,.-_aesni_ctr32_ghash_no_movbe_6x
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
.globl aesni_gcm_decrypt
.type aesni_gcm_decrypt,@function
.align 32
aesni_gcm_decrypt:
.cfi_startproc
xorq %r10,%r10
cmpq $0x60,%rdx
jb .Lgcm_dec_abort
leaq (%rsp),%rax
.cfi_def_cfa_register %rax
pushq %rbx
.cfi_offset %rbx,-16
pushq %rbp
.cfi_offset %rbp,-24
pushq %r12
.cfi_offset %r12,-32
pushq %r13
.cfi_offset %r13,-40
pushq %r14
.cfi_offset %r14,-48
pushq %r15
.cfi_offset %r15,-56
ICP: gcm: Allocate hash subkey table separately While evaluating other assembler implementations it turns out that the precomputed hash subkey tables vary in size, from 8*16 bytes (avx2/avx512) up to 48*16 bytes (avx512-vaes), depending on the implementation. To be able to handle the size differences later, allocate `gcm_Htable` dynamically rather then having a fixed size array, and adapt consumers. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11102
2020-10-30 22:24:21 +00:00
pushq %r9
.cfi_offset %r9,-64
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
vzeroupper
vmovdqu (%r8),%xmm1
addq $-128,%rsp
movl 12(%r8),%ebx
leaq .Lbswap_mask(%rip),%r11
leaq -128(%rcx),%r14
movq $0xf80,%r15
vmovdqu (%r9),%xmm8
andq $-128,%rsp
vmovdqu (%r11),%xmm0
leaq 128(%rcx),%rcx
ICP: gcm: Allocate hash subkey table separately While evaluating other assembler implementations it turns out that the precomputed hash subkey tables vary in size, from 8*16 bytes (avx2/avx512) up to 48*16 bytes (avx512-vaes), depending on the implementation. To be able to handle the size differences later, allocate `gcm_Htable` dynamically rather then having a fixed size array, and adapt consumers. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11102
2020-10-30 22:24:21 +00:00
movq 32(%r9),%r9
leaq 32(%r9),%r9
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
movl 504-128(%rcx),%ebp // ICP has a larger offset for rounds.
vpshufb %xmm0,%xmm8,%xmm8
andq %r15,%r14
andq %rsp,%r15
subq %r14,%r15
jc .Ldec_no_key_aliasing
cmpq $768,%r15
jnc .Ldec_no_key_aliasing
subq %r15,%rsp
.Ldec_no_key_aliasing:
vmovdqu 80(%rdi),%xmm7
leaq (%rdi),%r14
vmovdqu 64(%rdi),%xmm4
leaq -192(%rdi,%rdx,1),%r15
vmovdqu 48(%rdi),%xmm5
shrq $4,%rdx
xorq %r10,%r10
vmovdqu 32(%rdi),%xmm6
vpshufb %xmm0,%xmm7,%xmm7
vmovdqu 16(%rdi),%xmm2
vpshufb %xmm0,%xmm4,%xmm4
vmovdqu (%rdi),%xmm3
vpshufb %xmm0,%xmm5,%xmm5
vmovdqu %xmm4,48(%rsp)
vpshufb %xmm0,%xmm6,%xmm6
vmovdqu %xmm5,64(%rsp)
vpshufb %xmm0,%xmm2,%xmm2
vmovdqu %xmm6,80(%rsp)
vpshufb %xmm0,%xmm3,%xmm3
vmovdqu %xmm2,96(%rsp)
vmovdqu %xmm3,112(%rsp)
ICP: gcm-avx: Support architectures lacking the MOVBE instruction There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup #9749 Closes #10029
2020-03-17 17:24:38 +00:00
#ifdef HAVE_MOVBE
#ifdef _KERNEL
testl $1,gcm_avx_can_use_movbe(%rip)
#else
testl $1,gcm_avx_can_use_movbe@GOTPCREL(%rip)
#endif
jz 1f
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
call _aesni_ctr32_ghash_6x
ICP: gcm-avx: Support architectures lacking the MOVBE instruction There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup #9749 Closes #10029
2020-03-17 17:24:38 +00:00
jmp 2f
1:
#endif
call _aesni_ctr32_ghash_no_movbe_6x
2:
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
vmovups %xmm9,-96(%rsi)
vmovups %xmm10,-80(%rsi)
vmovups %xmm11,-64(%rsi)
vmovups %xmm12,-48(%rsi)
vmovups %xmm13,-32(%rsi)
vmovups %xmm14,-16(%rsi)
vpshufb (%r11),%xmm8,%xmm8
ICP: gcm: Allocate hash subkey table separately While evaluating other assembler implementations it turns out that the precomputed hash subkey tables vary in size, from 8*16 bytes (avx2/avx512) up to 48*16 bytes (avx512-vaes), depending on the implementation. To be able to handle the size differences later, allocate `gcm_Htable` dynamically rather then having a fixed size array, and adapt consumers. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11102
2020-10-30 22:24:21 +00:00
movq -56(%rax),%r9
.cfi_restore %r9
vmovdqu %xmm8,(%r9)
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
vzeroupper
movq -48(%rax),%r15
.cfi_restore %r15
movq -40(%rax),%r14
.cfi_restore %r14
movq -32(%rax),%r13
.cfi_restore %r13
movq -24(%rax),%r12
.cfi_restore %r12
movq -16(%rax),%rbp
.cfi_restore %rbp
movq -8(%rax),%rbx
.cfi_restore %rbx
leaq (%rax),%rsp
.cfi_def_cfa_register %rsp
.Lgcm_dec_abort:
movq %r10,%rax
.byte 0xf3,0xc3
.cfi_endproc
.size aesni_gcm_decrypt,.-aesni_gcm_decrypt
.type _aesni_ctr32_6x,@function
.align 32
_aesni_ctr32_6x:
Add some missing cfi frame info in aesni-gcm-x86_64.S While preparing #9749 some .cfi_{start,end}proc directives were missed. Add the missing ones. See upstream https://github.com/openssl/openssl/commit/275a048f Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11101
2020-10-30 22:23:18 +00:00
.cfi_startproc
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
vmovdqu 0-128(%rcx),%xmm4
vmovdqu 32(%r11),%xmm2
leaq -2(%rbp),%r13 // ICP uses 10,12,14 not 9,11,13 for rounds.
vmovups 16-128(%rcx),%xmm15
leaq 32-128(%rcx),%r12
vpxor %xmm4,%xmm1,%xmm9
addl $100663296,%ebx
jc .Lhandle_ctr32_2
vpaddb %xmm2,%xmm1,%xmm10
vpaddb %xmm2,%xmm10,%xmm11
vpxor %xmm4,%xmm10,%xmm10
vpaddb %xmm2,%xmm11,%xmm12
vpxor %xmm4,%xmm11,%xmm11
vpaddb %xmm2,%xmm12,%xmm13
vpxor %xmm4,%xmm12,%xmm12
vpaddb %xmm2,%xmm13,%xmm14
vpxor %xmm4,%xmm13,%xmm13
vpaddb %xmm2,%xmm14,%xmm1
vpxor %xmm4,%xmm14,%xmm14
jmp .Loop_ctr32
.align 16
.Loop_ctr32:
vaesenc %xmm15,%xmm9,%xmm9
vaesenc %xmm15,%xmm10,%xmm10
vaesenc %xmm15,%xmm11,%xmm11
vaesenc %xmm15,%xmm12,%xmm12
vaesenc %xmm15,%xmm13,%xmm13
vaesenc %xmm15,%xmm14,%xmm14
vmovups (%r12),%xmm15
leaq 16(%r12),%r12
decl %r13d
jnz .Loop_ctr32
vmovdqu (%r12),%xmm3
vaesenc %xmm15,%xmm9,%xmm9
vpxor 0(%rdi),%xmm3,%xmm4
vaesenc %xmm15,%xmm10,%xmm10
vpxor 16(%rdi),%xmm3,%xmm5
vaesenc %xmm15,%xmm11,%xmm11
vpxor 32(%rdi),%xmm3,%xmm6
vaesenc %xmm15,%xmm12,%xmm12
vpxor 48(%rdi),%xmm3,%xmm8
vaesenc %xmm15,%xmm13,%xmm13
vpxor 64(%rdi),%xmm3,%xmm2
vaesenc %xmm15,%xmm14,%xmm14
vpxor 80(%rdi),%xmm3,%xmm3
leaq 96(%rdi),%rdi
vaesenclast %xmm4,%xmm9,%xmm9
vaesenclast %xmm5,%xmm10,%xmm10
vaesenclast %xmm6,%xmm11,%xmm11
vaesenclast %xmm8,%xmm12,%xmm12
vaesenclast %xmm2,%xmm13,%xmm13
vaesenclast %xmm3,%xmm14,%xmm14
vmovups %xmm9,0(%rsi)
vmovups %xmm10,16(%rsi)
vmovups %xmm11,32(%rsi)
vmovups %xmm12,48(%rsi)
vmovups %xmm13,64(%rsi)
vmovups %xmm14,80(%rsi)
leaq 96(%rsi),%rsi
.byte 0xf3,0xc3
.align 32
.Lhandle_ctr32_2:
vpshufb %xmm0,%xmm1,%xmm6
vmovdqu 48(%r11),%xmm5
vpaddd 64(%r11),%xmm6,%xmm10
vpaddd %xmm5,%xmm6,%xmm11
vpaddd %xmm5,%xmm10,%xmm12
vpshufb %xmm0,%xmm10,%xmm10
vpaddd %xmm5,%xmm11,%xmm13
vpshufb %xmm0,%xmm11,%xmm11
vpxor %xmm4,%xmm10,%xmm10
vpaddd %xmm5,%xmm12,%xmm14
vpshufb %xmm0,%xmm12,%xmm12
vpxor %xmm4,%xmm11,%xmm11
vpaddd %xmm5,%xmm13,%xmm1
vpshufb %xmm0,%xmm13,%xmm13
vpxor %xmm4,%xmm12,%xmm12
vpshufb %xmm0,%xmm14,%xmm14
vpxor %xmm4,%xmm13,%xmm13
vpshufb %xmm0,%xmm1,%xmm1
vpxor %xmm4,%xmm14,%xmm14
jmp .Loop_ctr32
Add some missing cfi frame info in aesni-gcm-x86_64.S While preparing #9749 some .cfi_{start,end}proc directives were missed. Add the missing ones. See upstream https://github.com/openssl/openssl/commit/275a048f Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11101
2020-10-30 22:23:18 +00:00
.cfi_endproc
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
.size _aesni_ctr32_6x,.-_aesni_ctr32_6x
.globl aesni_gcm_encrypt
.type aesni_gcm_encrypt,@function
.align 32
aesni_gcm_encrypt:
.cfi_startproc
xorq %r10,%r10
cmpq $288,%rdx
jb .Lgcm_enc_abort
leaq (%rsp),%rax
.cfi_def_cfa_register %rax
pushq %rbx
.cfi_offset %rbx,-16
pushq %rbp
.cfi_offset %rbp,-24
pushq %r12
.cfi_offset %r12,-32
pushq %r13
.cfi_offset %r13,-40
pushq %r14
.cfi_offset %r14,-48
pushq %r15
.cfi_offset %r15,-56
ICP: gcm: Allocate hash subkey table separately While evaluating other assembler implementations it turns out that the precomputed hash subkey tables vary in size, from 8*16 bytes (avx2/avx512) up to 48*16 bytes (avx512-vaes), depending on the implementation. To be able to handle the size differences later, allocate `gcm_Htable` dynamically rather then having a fixed size array, and adapt consumers. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11102
2020-10-30 22:24:21 +00:00
pushq %r9
.cfi_offset %r9,-64
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
vzeroupper
vmovdqu (%r8),%xmm1
addq $-128,%rsp
movl 12(%r8),%ebx
leaq .Lbswap_mask(%rip),%r11
leaq -128(%rcx),%r14
movq $0xf80,%r15
leaq 128(%rcx),%rcx
vmovdqu (%r11),%xmm0
andq $-128,%rsp
movl 504-128(%rcx),%ebp // ICP has an larger offset for rounds.
andq %r15,%r14
andq %rsp,%r15
subq %r14,%r15
jc .Lenc_no_key_aliasing
cmpq $768,%r15
jnc .Lenc_no_key_aliasing
subq %r15,%rsp
.Lenc_no_key_aliasing:
leaq (%rsi),%r14
leaq -192(%rsi,%rdx,1),%r15
shrq $4,%rdx
call _aesni_ctr32_6x
vpshufb %xmm0,%xmm9,%xmm8
vpshufb %xmm0,%xmm10,%xmm2
vmovdqu %xmm8,112(%rsp)
vpshufb %xmm0,%xmm11,%xmm4
vmovdqu %xmm2,96(%rsp)
vpshufb %xmm0,%xmm12,%xmm5
vmovdqu %xmm4,80(%rsp)
vpshufb %xmm0,%xmm13,%xmm6
vmovdqu %xmm5,64(%rsp)
vpshufb %xmm0,%xmm14,%xmm7
vmovdqu %xmm6,48(%rsp)
call _aesni_ctr32_6x
vmovdqu (%r9),%xmm8
ICP: gcm: Allocate hash subkey table separately While evaluating other assembler implementations it turns out that the precomputed hash subkey tables vary in size, from 8*16 bytes (avx2/avx512) up to 48*16 bytes (avx512-vaes), depending on the implementation. To be able to handle the size differences later, allocate `gcm_Htable` dynamically rather then having a fixed size array, and adapt consumers. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11102
2020-10-30 22:24:21 +00:00
movq 32(%r9),%r9
leaq 32(%r9),%r9
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
subq $12,%rdx
movq $192,%r10
vpshufb %xmm0,%xmm8,%xmm8
ICP: gcm-avx: Support architectures lacking the MOVBE instruction There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup #9749 Closes #10029
2020-03-17 17:24:38 +00:00
#ifdef HAVE_MOVBE
#ifdef _KERNEL
testl $1,gcm_avx_can_use_movbe(%rip)
#else
testl $1,gcm_avx_can_use_movbe@GOTPCREL(%rip)
#endif
jz 1f
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
call _aesni_ctr32_ghash_6x
ICP: gcm-avx: Support architectures lacking the MOVBE instruction There are a couple of x86_64 architectures which support all needed features to make the accelerated GCM implementation work but the MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge and AMD Bulldozer, Piledriver, and Steamroller. By using MOVBE only if available and replacing it with a MOV followed by a BSWAP if not, those architectures now benefit from the new GCM routines and performance is considerably better compared to the original implementation. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-by: Adam D. Moss <c@yotes.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Followup #9749 Closes #10029
2020-03-17 17:24:38 +00:00
jmp 2f
1:
#endif
call _aesni_ctr32_ghash_no_movbe_6x
2:
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
vmovdqu 32(%rsp),%xmm7
vmovdqu (%r11),%xmm0
vmovdqu 0-32(%r9),%xmm3
vpunpckhqdq %xmm7,%xmm7,%xmm1
vmovdqu 32-32(%r9),%xmm15
vmovups %xmm9,-96(%rsi)
vpshufb %xmm0,%xmm9,%xmm9
vpxor %xmm7,%xmm1,%xmm1
vmovups %xmm10,-80(%rsi)
vpshufb %xmm0,%xmm10,%xmm10
vmovups %xmm11,-64(%rsi)
vpshufb %xmm0,%xmm11,%xmm11
vmovups %xmm12,-48(%rsi)
vpshufb %xmm0,%xmm12,%xmm12
vmovups %xmm13,-32(%rsi)
vpshufb %xmm0,%xmm13,%xmm13
vmovups %xmm14,-16(%rsi)
vpshufb %xmm0,%xmm14,%xmm14
vmovdqu %xmm9,16(%rsp)
vmovdqu 48(%rsp),%xmm6
vmovdqu 16-32(%r9),%xmm0
vpunpckhqdq %xmm6,%xmm6,%xmm2
vpclmulqdq $0x00,%xmm3,%xmm7,%xmm5
vpxor %xmm6,%xmm2,%xmm2
vpclmulqdq $0x11,%xmm3,%xmm7,%xmm7
vpclmulqdq $0x00,%xmm15,%xmm1,%xmm1
vmovdqu 64(%rsp),%xmm9
vpclmulqdq $0x00,%xmm0,%xmm6,%xmm4
vmovdqu 48-32(%r9),%xmm3
vpxor %xmm5,%xmm4,%xmm4
vpunpckhqdq %xmm9,%xmm9,%xmm5
vpclmulqdq $0x11,%xmm0,%xmm6,%xmm6
vpxor %xmm9,%xmm5,%xmm5
vpxor %xmm7,%xmm6,%xmm6
vpclmulqdq $0x10,%xmm15,%xmm2,%xmm2
vmovdqu 80-32(%r9),%xmm15
vpxor %xmm1,%xmm2,%xmm2
vmovdqu 80(%rsp),%xmm1
vpclmulqdq $0x00,%xmm3,%xmm9,%xmm7
vmovdqu 64-32(%r9),%xmm0
vpxor %xmm4,%xmm7,%xmm7
vpunpckhqdq %xmm1,%xmm1,%xmm4
vpclmulqdq $0x11,%xmm3,%xmm9,%xmm9
vpxor %xmm1,%xmm4,%xmm4
vpxor %xmm6,%xmm9,%xmm9
vpclmulqdq $0x00,%xmm15,%xmm5,%xmm5
vpxor %xmm2,%xmm5,%xmm5
vmovdqu 96(%rsp),%xmm2
vpclmulqdq $0x00,%xmm0,%xmm1,%xmm6
vmovdqu 96-32(%r9),%xmm3
vpxor %xmm7,%xmm6,%xmm6
vpunpckhqdq %xmm2,%xmm2,%xmm7
vpclmulqdq $0x11,%xmm0,%xmm1,%xmm1
vpxor %xmm2,%xmm7,%xmm7
vpxor %xmm9,%xmm1,%xmm1
vpclmulqdq $0x10,%xmm15,%xmm4,%xmm4
vmovdqu 128-32(%r9),%xmm15
vpxor %xmm5,%xmm4,%xmm4
vpxor 112(%rsp),%xmm8,%xmm8
vpclmulqdq $0x00,%xmm3,%xmm2,%xmm5
vmovdqu 112-32(%r9),%xmm0
vpunpckhqdq %xmm8,%xmm8,%xmm9
vpxor %xmm6,%xmm5,%xmm5
vpclmulqdq $0x11,%xmm3,%xmm2,%xmm2
vpxor %xmm8,%xmm9,%xmm9
vpxor %xmm1,%xmm2,%xmm2
vpclmulqdq $0x00,%xmm15,%xmm7,%xmm7
vpxor %xmm4,%xmm7,%xmm4
vpclmulqdq $0x00,%xmm0,%xmm8,%xmm6
vmovdqu 0-32(%r9),%xmm3
vpunpckhqdq %xmm14,%xmm14,%xmm1
vpclmulqdq $0x11,%xmm0,%xmm8,%xmm8
vpxor %xmm14,%xmm1,%xmm1
vpxor %xmm5,%xmm6,%xmm5
vpclmulqdq $0x10,%xmm15,%xmm9,%xmm9
vmovdqu 32-32(%r9),%xmm15
vpxor %xmm2,%xmm8,%xmm7
vpxor %xmm4,%xmm9,%xmm6
vmovdqu 16-32(%r9),%xmm0
vpxor %xmm5,%xmm7,%xmm9
vpclmulqdq $0x00,%xmm3,%xmm14,%xmm4
vpxor %xmm9,%xmm6,%xmm6
vpunpckhqdq %xmm13,%xmm13,%xmm2
vpclmulqdq $0x11,%xmm3,%xmm14,%xmm14
vpxor %xmm13,%xmm2,%xmm2
vpslldq $8,%xmm6,%xmm9
vpclmulqdq $0x00,%xmm15,%xmm1,%xmm1
vpxor %xmm9,%xmm5,%xmm8
vpsrldq $8,%xmm6,%xmm6
vpxor %xmm6,%xmm7,%xmm7
vpclmulqdq $0x00,%xmm0,%xmm13,%xmm5
vmovdqu 48-32(%r9),%xmm3
vpxor %xmm4,%xmm5,%xmm5
vpunpckhqdq %xmm12,%xmm12,%xmm9
vpclmulqdq $0x11,%xmm0,%xmm13,%xmm13
vpxor %xmm12,%xmm9,%xmm9
vpxor %xmm14,%xmm13,%xmm13
vpalignr $8,%xmm8,%xmm8,%xmm14
vpclmulqdq $0x10,%xmm15,%xmm2,%xmm2
vmovdqu 80-32(%r9),%xmm15
vpxor %xmm1,%xmm2,%xmm2
vpclmulqdq $0x00,%xmm3,%xmm12,%xmm4
vmovdqu 64-32(%r9),%xmm0
vpxor %xmm5,%xmm4,%xmm4
vpunpckhqdq %xmm11,%xmm11,%xmm1
vpclmulqdq $0x11,%xmm3,%xmm12,%xmm12
vpxor %xmm11,%xmm1,%xmm1
vpxor %xmm13,%xmm12,%xmm12
vxorps 16(%rsp),%xmm7,%xmm7
vpclmulqdq $0x00,%xmm15,%xmm9,%xmm9
vpxor %xmm2,%xmm9,%xmm9
vpclmulqdq $0x10,16(%r11),%xmm8,%xmm8
vxorps %xmm14,%xmm8,%xmm8
vpclmulqdq $0x00,%xmm0,%xmm11,%xmm5
vmovdqu 96-32(%r9),%xmm3
vpxor %xmm4,%xmm5,%xmm5
vpunpckhqdq %xmm10,%xmm10,%xmm2
vpclmulqdq $0x11,%xmm0,%xmm11,%xmm11
vpxor %xmm10,%xmm2,%xmm2
vpalignr $8,%xmm8,%xmm8,%xmm14
vpxor %xmm12,%xmm11,%xmm11
vpclmulqdq $0x10,%xmm15,%xmm1,%xmm1
vmovdqu 128-32(%r9),%xmm15
vpxor %xmm9,%xmm1,%xmm1
vxorps %xmm7,%xmm14,%xmm14
vpclmulqdq $0x10,16(%r11),%xmm8,%xmm8
vxorps %xmm14,%xmm8,%xmm8
vpclmulqdq $0x00,%xmm3,%xmm10,%xmm4
vmovdqu 112-32(%r9),%xmm0
vpxor %xmm5,%xmm4,%xmm4
vpunpckhqdq %xmm8,%xmm8,%xmm9
vpclmulqdq $0x11,%xmm3,%xmm10,%xmm10
vpxor %xmm8,%xmm9,%xmm9
vpxor %xmm11,%xmm10,%xmm10
vpclmulqdq $0x00,%xmm15,%xmm2,%xmm2
vpxor %xmm1,%xmm2,%xmm2
vpclmulqdq $0x00,%xmm0,%xmm8,%xmm5
vpclmulqdq $0x11,%xmm0,%xmm8,%xmm7
vpxor %xmm4,%xmm5,%xmm5
vpclmulqdq $0x10,%xmm15,%xmm9,%xmm6
vpxor %xmm10,%xmm7,%xmm7
vpxor %xmm2,%xmm6,%xmm6
vpxor %xmm5,%xmm7,%xmm4
vpxor %xmm4,%xmm6,%xmm6
vpslldq $8,%xmm6,%xmm1
vmovdqu 16(%r11),%xmm3
vpsrldq $8,%xmm6,%xmm6
vpxor %xmm1,%xmm5,%xmm8
vpxor %xmm6,%xmm7,%xmm7
vpalignr $8,%xmm8,%xmm8,%xmm2
vpclmulqdq $0x10,%xmm3,%xmm8,%xmm8
vpxor %xmm2,%xmm8,%xmm8
vpalignr $8,%xmm8,%xmm8,%xmm2
vpclmulqdq $0x10,%xmm3,%xmm8,%xmm8
vpxor %xmm7,%xmm2,%xmm2
vpxor %xmm2,%xmm8,%xmm8
vpshufb (%r11),%xmm8,%xmm8
ICP: gcm: Allocate hash subkey table separately While evaluating other assembler implementations it turns out that the precomputed hash subkey tables vary in size, from 8*16 bytes (avx2/avx512) up to 48*16 bytes (avx512-vaes), depending on the implementation. To be able to handle the size differences later, allocate `gcm_Htable` dynamically rather then having a fixed size array, and adapt consumers. Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #11102
2020-10-30 22:24:21 +00:00
movq -56(%rax),%r9
.cfi_restore %r9
vmovdqu %xmm8,(%r9)
ICP: Improve AES-GCM performance Currently SIMD accelerated AES-GCM performance is limited by two factors: a. The need to disable preemption and interrupts and save the FPU state before using it and to do the reverse when done. Due to the way the code is organized (see (b) below) we have to pay this price twice for each 16 byte GCM block processed. b. Most processing is done in C, operating on single GCM blocks. The use of SIMD instructions is limited to the AES encryption of the counter block (AES-NI) and the Galois multiplication (PCLMULQDQ). This leads to the FPU not being fully utilized for crypto operations. To solve (a) we do crypto processing in larger chunks while owning the FPU. An `icp_gcm_avx_chunk_size` module parameter was introduced to make this chunk size tweakable. It defaults to 32 KiB. This step alone roughly doubles performance. (b) is tackled by porting and using the highly optimized openssl AES-GCM assembler routines, which do all the processing (CTR, AES, GMULT) in a single routine. Both steps together result in up to 32x reduction of the time spend in the en/decryption routines, leading up to approximately 12x throughput increase for large (128 KiB) blocks. Lastly, this commit changes the default encryption algorithm from AES-CCM to AES-GCM when setting the `encryption=on` property. Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov> Reviewed-By: Jason King <jason.king@joyent.com> Reviewed-By: Tom Caputi <tcaputi@datto.com> Reviewed-By: Richard Laager <rlaager@wiktel.com> Signed-off-by: Attila Fülöp <attila@fueloep.org> Closes #9749
2020-02-10 20:59:50 +00:00
vzeroupper
movq -48(%rax),%r15
.cfi_restore %r15
movq -40(%rax),%r14
.cfi_restore %r14
movq -32(%rax),%r13
.cfi_restore %r13
movq -24(%rax),%r12
.cfi_restore %r12
movq -16(%rax),%rbp
.cfi_restore %rbp
movq -8(%rax),%rbx
.cfi_restore %rbx
leaq (%rax),%rsp
.cfi_def_cfa_register %rsp
.Lgcm_enc_abort:
movq %r10,%rax
.byte 0xf3,0xc3
.cfi_endproc
.size aesni_gcm_encrypt,.-aesni_gcm_encrypt
/* Some utility routines */
/*
* clear all fpu registers
* void clear_fpu_regs_avx(void);
*/
.globl clear_fpu_regs_avx
.type clear_fpu_regs_avx,@function
.align 32
clear_fpu_regs_avx:
vzeroall
ret
.size clear_fpu_regs_avx,.-clear_fpu_regs_avx
/*
* void gcm_xor_avx(const uint8_t *src, uint8_t *dst);
*
* XORs one pair of unaligned 128-bit blocks from `src' and `dst' and
* stores the result at `dst'. The XOR is performed using FPU registers,
* so make sure FPU state is saved when running this in the kernel.
*/
.globl gcm_xor_avx
.type gcm_xor_avx,@function
.align 32
gcm_xor_avx:
movdqu (%rdi), %xmm0
movdqu (%rsi), %xmm1
pxor %xmm1, %xmm0
movdqu %xmm0, (%rsi)
ret
.size gcm_xor_avx,.-gcm_xor_avx
/*
* Toggle a boolean_t value atomically and return the new value.
* boolean_t atomic_toggle_boolean_nv(volatile boolean_t *);
*/
.globl atomic_toggle_boolean_nv
.type atomic_toggle_boolean_nv,@function
.align 32
atomic_toggle_boolean_nv:
xorl %eax, %eax
lock
xorl $1, (%rdi)
jz 1f
movl $1, %eax
1:
ret
.size atomic_toggle_boolean_nv,.-atomic_toggle_boolean_nv
.align 64
.Lbswap_mask:
.byte 15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
.Lpoly:
.byte 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
.Lone_msb:
.byte 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1
.Ltwo_lsb:
.byte 2,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
.Lone_lsb:
.byte 1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
.byte 65,69,83,45,78,73,32,71,67,77,32,109,111,100,117,108,101,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
.align 64
/* Mark the stack non-executable. */
#if defined(__linux__) && defined(__ELF__)
.section .note.GNU-stack,"",%progbits
#endif
#endif /* defined(__x86_64__) && defined(HAVE_AVX) && defined(HAVE_AES) ... */