Skip to content

Commit 0f961f9

Browse files
ebiggersherbertx
ebiggers
authored and
herbertx
committed
crypto: x86/nhpoly1305 - add AVX2 accelerated NHPoly1305
Add a 64-bit AVX2 implementation of NHPoly1305, an ε-almost-∆-universal hash function used in the Adiantum encryption mode. For now, only the NH portion is actually AVX2-accelerated; the Poly1305 part is less performance-critical so is just implemented in C. Signed-off-by: Eric Biggers <[email protected]> Signed-off-by: Herbert Xu <[email protected]>
1 parent 012c823 commit 0f961f9

File tree

4 files changed

+245
-0
lines changed

4 files changed

+245
-0
lines changed

‎arch/x86/crypto/Makefile‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -48,6 +48,7 @@ obj-$(CONFIG_CRYPTO_MORUS640_SSE2) += morus640-sse2.o
4848
obj-$(CONFIG_CRYPTO_MORUS1280_SSE2) += morus1280-sse2.o
4949

5050
obj-$(CONFIG_CRYPTO_NHPOLY1305_SSE2) += nhpoly1305-sse2.o
51+
obj-$(CONFIG_CRYPTO_NHPOLY1305_AVX2) += nhpoly1305-avx2.o
5152

5253
# These modules require assembler to support AVX.
5354
ifeq ($(avx_supported),yes)
@@ -106,6 +107,8 @@ ifeq ($(avx2_supported),yes)
106107
serpent-avx2-y := serpent-avx2-asm_64.o serpent_avx2_glue.o
107108

108109
morus1280-avx2-y := morus1280-avx2-asm.o morus1280-avx2-glue.o
110+
111+
nhpoly1305-avx2-y := nh-avx2-x86_64.o nhpoly1305-avx2-glue.o
109112
endif
110113

111114
ifeq ($(avx512_supported),yes)

‎arch/x86/crypto/nh-avx2-x86_64.S‎

Lines changed: 157 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,157 @@
1+
/* SPDX-License-Identifier: GPL-2.0 */
2+
/*
3+
* NH - ε-almost-universal hash function, x86_64 AVX2 accelerated
4+
*
5+
* Copyright 2018 Google LLC
6+
*
7+
* Author: Eric Biggers <[email protected]>
8+
*/
9+
10+
#include <linux/linkage.h>
11+
12+
#define PASS0_SUMS %ymm0
13+
#define PASS1_SUMS %ymm1
14+
#define PASS2_SUMS %ymm2
15+
#define PASS3_SUMS %ymm3
16+
#define K0 %ymm4
17+
#define K0_XMM %xmm4
18+
#define K1 %ymm5
19+
#define K1_XMM %xmm5
20+
#define K2 %ymm6
21+
#define K2_XMM %xmm6
22+
#define K3 %ymm7
23+
#define K3_XMM %xmm7
24+
#define T0 %ymm8
25+
#define T1 %ymm9
26+
#define T2 %ymm10
27+
#define T2_XMM %xmm10
28+
#define T3 %ymm11
29+
#define T3_XMM %xmm11
30+
#define T4 %ymm12
31+
#define T5 %ymm13
32+
#define T6 %ymm14
33+
#define T7 %ymm15
34+
#define KEY %rdi
35+
#define MESSAGE %rsi
36+
#define MESSAGE_LEN %rdx
37+
#define HASH %rcx
38+
39+
.macro _nh_2xstride k0, k1, k2, k3
40+
41+
// Add message words to key words
42+
vpaddd \k0, T3, T0
43+
vpaddd \k1, T3, T1
44+
vpaddd \k2, T3, T2
45+
vpaddd \k3, T3, T3
46+
47+
// Multiply 32x32 => 64 and accumulate
48+
vpshufd $0x10, T0, T4
49+
vpshufd $0x32, T0, T0
50+
vpshufd $0x10, T1, T5
51+
vpshufd $0x32, T1, T1
52+
vpshufd $0x10, T2, T6
53+
vpshufd $0x32, T2, T2
54+
vpshufd $0x10, T3, T7
55+
vpshufd $0x32, T3, T3
56+
vpmuludq T4, T0, T0
57+
vpmuludq T5, T1, T1
58+
vpmuludq T6, T2, T2
59+
vpmuludq T7, T3, T3
60+
vpaddq T0, PASS0_SUMS, PASS0_SUMS
61+
vpaddq T1, PASS1_SUMS, PASS1_SUMS
62+
vpaddq T2, PASS2_SUMS, PASS2_SUMS
63+
vpaddq T3, PASS3_SUMS, PASS3_SUMS
64+
.endm
65+
66+
/*
67+
* void nh_avx2(const u32 *key, const u8 *message, size_t message_len,
68+
* u8 hash[NH_HASH_BYTES])
69+
*
70+
* It's guaranteed that message_len % 16 == 0.
71+
*/
72+
ENTRY(nh_avx2)
73+
74+
vmovdqu 0x00(KEY), K0
75+
vmovdqu 0x10(KEY), K1
76+
add $0x20, KEY
77+
vpxor PASS0_SUMS, PASS0_SUMS, PASS0_SUMS
78+
vpxor PASS1_SUMS, PASS1_SUMS, PASS1_SUMS
79+
vpxor PASS2_SUMS, PASS2_SUMS, PASS2_SUMS
80+
vpxor PASS3_SUMS, PASS3_SUMS, PASS3_SUMS
81+
82+
sub $0x40, MESSAGE_LEN
83+
jl .Lloop4_done
84+
.Lloop4:
85+
vmovdqu (MESSAGE), T3
86+
vmovdqu 0x00(KEY), K2
87+
vmovdqu 0x10(KEY), K3
88+
_nh_2xstride K0, K1, K2, K3
89+
90+
vmovdqu 0x20(MESSAGE), T3
91+
vmovdqu 0x20(KEY), K0
92+
vmovdqu 0x30(KEY), K1
93+
_nh_2xstride K2, K3, K0, K1
94+
95+
add $0x40, MESSAGE
96+
add $0x40, KEY
97+
sub $0x40, MESSAGE_LEN
98+
jge .Lloop4
99+
100+
.Lloop4_done:
101+
and $0x3f, MESSAGE_LEN
102+
jz .Ldone
103+
104+
cmp $0x20, MESSAGE_LEN
105+
jl .Llast
106+
107+
// 2 or 3 strides remain; do 2 more.
108+
vmovdqu (MESSAGE), T3
109+
vmovdqu 0x00(KEY), K2
110+
vmovdqu 0x10(KEY), K3
111+
_nh_2xstride K0, K1, K2, K3
112+
add $0x20, MESSAGE
113+
add $0x20, KEY
114+
sub $0x20, MESSAGE_LEN
115+
jz .Ldone
116+
vmovdqa K2, K0
117+
vmovdqa K3, K1
118+
.Llast:
119+
// Last stride. Zero the high 128 bits of the message and keys so they
120+
// don't affect the result when processing them like 2 strides.
121+
vmovdqu (MESSAGE), T3_XMM
122+
vmovdqa K0_XMM, K0_XMM
123+
vmovdqa K1_XMM, K1_XMM
124+
vmovdqu 0x00(KEY), K2_XMM
125+
vmovdqu 0x10(KEY), K3_XMM
126+
_nh_2xstride K0, K1, K2, K3
127+
128+
.Ldone:
129+
// Sum the accumulators for each pass, then store the sums to 'hash'
130+
131+
// PASS0_SUMS is (0A 0B 0C 0D)
132+
// PASS1_SUMS is (1A 1B 1C 1D)
133+
// PASS2_SUMS is (2A 2B 2C 2D)
134+
// PASS3_SUMS is (3A 3B 3C 3D)
135+
// We need the horizontal sums:
136+
// (0A + 0B + 0C + 0D,
137+
// 1A + 1B + 1C + 1D,
138+
// 2A + 2B + 2C + 2D,
139+
// 3A + 3B + 3C + 3D)
140+
//
141+
142+
vpunpcklqdq PASS1_SUMS, PASS0_SUMS, T0 // T0 = (0A 1A 0C 1C)
143+
vpunpckhqdq PASS1_SUMS, PASS0_SUMS, T1 // T1 = (0B 1B 0D 1D)
144+
vpunpcklqdq PASS3_SUMS, PASS2_SUMS, T2 // T2 = (2A 3A 2C 3C)
145+
vpunpckhqdq PASS3_SUMS, PASS2_SUMS, T3 // T3 = (2B 3B 2D 3D)
146+
147+
vinserti128 $0x1, T2_XMM, T0, T4 // T4 = (0A 1A 2A 3A)
148+
vinserti128 $0x1, T3_XMM, T1, T5 // T5 = (0B 1B 2B 3B)
149+
vperm2i128 $0x31, T2, T0, T0 // T0 = (0C 1C 2C 3C)
150+
vperm2i128 $0x31, T3, T1, T1 // T1 = (0D 1D 2D 3D)
151+
152+
vpaddq T5, T4, T4
153+
vpaddq T1, T0, T0
154+
vpaddq T4, T0, T0
155+
vmovdqu T0, (HASH)
156+
ret
157+
ENDPROC(nh_avx2)

‎arch/x86/crypto/nhpoly1305-avx2-glue.c‎

Lines changed: 77 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,77 @@
1+
// SPDX-License-Identifier: GPL-2.0
2+
/*
3+
* NHPoly1305 - ε-almost-∆-universal hash function for Adiantum
4+
* (AVX2 accelerated version)
5+
*
6+
* Copyright 2018 Google LLC
7+
*/
8+
9+
#include <crypto/internal/hash.h>
10+
#include <crypto/nhpoly1305.h>
11+
#include <linux/module.h>
12+
#include <asm/fpu/api.h>
13+
14+
asmlinkage void nh_avx2(const u32 *key, const u8 *message, size_t message_len,
15+
u8 hash[NH_HASH_BYTES]);
16+
17+
/* wrapper to avoid indirect call to assembly, which doesn't work with CFI */
18+
static void _nh_avx2(const u32 *key, const u8 *message, size_t message_len,
19+
__le64 hash[NH_NUM_PASSES])
20+
{
21+
nh_avx2(key, message, message_len, (u8 *)hash);
22+
}
23+
24+
static int nhpoly1305_avx2_update(struct shash_desc *desc,
25+
const u8 *src, unsigned int srclen)
26+
{
27+
if (srclen < 64 || !irq_fpu_usable())
28+
return crypto_nhpoly1305_update(desc, src, srclen);
29+
30+
do {
31+
unsigned int n = min_t(unsigned int, srclen, PAGE_SIZE);
32+
33+
kernel_fpu_begin();
34+
crypto_nhpoly1305_update_helper(desc, src, n, _nh_avx2);
35+
kernel_fpu_end();
36+
src += n;
37+
srclen -= n;
38+
} while (srclen);
39+
return 0;
40+
}
41+
42+
static struct shash_alg nhpoly1305_alg = {
43+
.base.cra_name = "nhpoly1305",
44+
.base.cra_driver_name = "nhpoly1305-avx2",
45+
.base.cra_priority = 300,
46+
.base.cra_ctxsize = sizeof(struct nhpoly1305_key),
47+
.base.cra_module = THIS_MODULE,
48+
.digestsize = POLY1305_DIGEST_SIZE,
49+
.init = crypto_nhpoly1305_init,
50+
.update = nhpoly1305_avx2_update,
51+
.final = crypto_nhpoly1305_final,
52+
.setkey = crypto_nhpoly1305_setkey,
53+
.descsize = sizeof(struct nhpoly1305_state),
54+
};
55+
56+
static int __init nhpoly1305_mod_init(void)
57+
{
58+
if (!boot_cpu_has(X86_FEATURE_AVX2) ||
59+
!boot_cpu_has(X86_FEATURE_OSXSAVE))
60+
return -ENODEV;
61+
62+
return crypto_register_shash(&nhpoly1305_alg);
63+
}
64+
65+
static void __exit nhpoly1305_mod_exit(void)
66+
{
67+
crypto_unregister_shash(&nhpoly1305_alg);
68+
}
69+
70+
module_init(nhpoly1305_mod_init);
71+
module_exit(nhpoly1305_mod_exit);
72+
73+
MODULE_DESCRIPTION("NHPoly1305 ε-almost-∆-universal hash function (AVX2-accelerated)");
74+
MODULE_LICENSE("GPL v2");
75+
MODULE_AUTHOR("Eric Biggers <[email protected]>");
76+
MODULE_ALIAS_CRYPTO("nhpoly1305");
77+
MODULE_ALIAS_CRYPTO("nhpoly1305-avx2");

‎crypto/Kconfig‎

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -509,6 +509,14 @@ config CRYPTO_NHPOLY1305_SSE2
509509
SSE2 optimized implementation of the hash function used by the
510510
Adiantum encryption mode.
511511

512+
config CRYPTO_NHPOLY1305_AVX2
513+
tristate "NHPoly1305 hash function (x86_64 AVX2 implementation)"
514+
depends on X86 && 64BIT
515+
select CRYPTO_NHPOLY1305
516+
help
517+
AVX2 optimized implementation of the hash function used by the
518+
Adiantum encryption mode.
519+
512520
config CRYPTO_ADIANTUM
513521
tristate "Adiantum support"
514522
select CRYPTO_CHACHA20

0 commit comments

Comments
 (0)