Book a Demo!
CoCalc Logo Icon
StoreFeaturesDocsShareSupportNewsAboutPoliciesSign UpSign In
torvalds
GitHub Repository: torvalds/linux
Path: blob/master/lib/crypto/x86/poly1305_glue.c
26292 views
1
// SPDX-License-Identifier: GPL-2.0 OR MIT
2
/*
3
* Copyright (C) 2015-2019 Jason A. Donenfeld <[email protected]>. All Rights Reserved.
4
*/
5
6
#include <asm/cpu_device_id.h>
7
#include <asm/fpu/api.h>
8
#include <crypto/internal/poly1305.h>
9
#include <linux/jump_label.h>
10
#include <linux/kernel.h>
11
#include <linux/module.h>
12
#include <linux/sizes.h>
13
#include <linux/unaligned.h>
14
15
struct poly1305_arch_internal {
16
union {
17
struct {
18
u32 h[5];
19
u32 is_base2_26;
20
};
21
u64 hs[3];
22
};
23
u64 r[2];
24
u64 pad;
25
struct { u32 r2, r1, r4, r3; } rn[9];
26
};
27
28
/*
29
* The AVX code uses base 2^26, while the scalar code uses base 2^64. If we hit
30
* the unfortunate situation of using AVX and then having to go back to scalar
31
* -- because the user is silly and has called the update function from two
32
* separate contexts -- then we need to convert back to the original base before
33
* proceeding. It is possible to reason that the initial reduction below is
34
* sufficient given the implementation invariants. However, for an avoidance of
35
* doubt and because this is not performance critical, we do the full reduction
36
* anyway. Z3 proof of below function: https://xn--4db.cc/ltPtHCKN/py
37
*/
38
static void convert_to_base2_64(void *ctx)
39
{
40
struct poly1305_arch_internal *state = ctx;
41
u32 cy;
42
43
if (!state->is_base2_26)
44
return;
45
46
cy = state->h[0] >> 26; state->h[0] &= 0x3ffffff; state->h[1] += cy;
47
cy = state->h[1] >> 26; state->h[1] &= 0x3ffffff; state->h[2] += cy;
48
cy = state->h[2] >> 26; state->h[2] &= 0x3ffffff; state->h[3] += cy;
49
cy = state->h[3] >> 26; state->h[3] &= 0x3ffffff; state->h[4] += cy;
50
state->hs[0] = ((u64)state->h[2] << 52) | ((u64)state->h[1] << 26) | state->h[0];
51
state->hs[1] = ((u64)state->h[4] << 40) | ((u64)state->h[3] << 14) | (state->h[2] >> 12);
52
state->hs[2] = state->h[4] >> 24;
53
/* Unsigned Less Than: branchlessly produces 1 if a < b, else 0. */
54
#define ULT(a, b) ((a ^ ((a ^ b) | ((a - b) ^ b))) >> (sizeof(a) * 8 - 1))
55
cy = (state->hs[2] >> 2) + (state->hs[2] & ~3ULL);
56
state->hs[2] &= 3;
57
state->hs[0] += cy;
58
state->hs[1] += (cy = ULT(state->hs[0], cy));
59
state->hs[2] += ULT(state->hs[1], cy);
60
#undef ULT
61
state->is_base2_26 = 0;
62
}
63
64
asmlinkage void poly1305_block_init_arch(
65
struct poly1305_block_state *state,
66
const u8 raw_key[POLY1305_BLOCK_SIZE]);
67
EXPORT_SYMBOL_GPL(poly1305_block_init_arch);
68
asmlinkage void poly1305_blocks_x86_64(struct poly1305_arch_internal *ctx,
69
const u8 *inp,
70
const size_t len, const u32 padbit);
71
asmlinkage void poly1305_emit_x86_64(const struct poly1305_state *ctx,
72
u8 mac[POLY1305_DIGEST_SIZE],
73
const u32 nonce[4]);
74
asmlinkage void poly1305_emit_avx(const struct poly1305_state *ctx,
75
u8 mac[POLY1305_DIGEST_SIZE],
76
const u32 nonce[4]);
77
asmlinkage void poly1305_blocks_avx(struct poly1305_arch_internal *ctx,
78
const u8 *inp, const size_t len,
79
const u32 padbit);
80
asmlinkage void poly1305_blocks_avx2(struct poly1305_arch_internal *ctx,
81
const u8 *inp, const size_t len,
82
const u32 padbit);
83
asmlinkage void poly1305_blocks_avx512(struct poly1305_arch_internal *ctx,
84
const u8 *inp,
85
const size_t len, const u32 padbit);
86
87
static __ro_after_init DEFINE_STATIC_KEY_FALSE(poly1305_use_avx);
88
static __ro_after_init DEFINE_STATIC_KEY_FALSE(poly1305_use_avx2);
89
static __ro_after_init DEFINE_STATIC_KEY_FALSE(poly1305_use_avx512);
90
91
void poly1305_blocks_arch(struct poly1305_block_state *state, const u8 *inp,
92
unsigned int len, u32 padbit)
93
{
94
struct poly1305_arch_internal *ctx =
95
container_of(&state->h.h, struct poly1305_arch_internal, h);
96
97
/* SIMD disables preemption, so relax after processing each page. */
98
BUILD_BUG_ON(SZ_4K < POLY1305_BLOCK_SIZE ||
99
SZ_4K % POLY1305_BLOCK_SIZE);
100
101
/*
102
* The AVX implementations have significant setup overhead (e.g. key
103
* power computation, kernel FPU enabling) which makes them slower for
104
* short messages. Fall back to the scalar implementation for messages
105
* shorter than 288 bytes, unless the AVX-specific key setup has already
106
* been performed (indicated by ctx->is_base2_26).
107
*/
108
if (!static_branch_likely(&poly1305_use_avx) ||
109
(len < POLY1305_BLOCK_SIZE * 18 && !ctx->is_base2_26) ||
110
unlikely(!irq_fpu_usable())) {
111
convert_to_base2_64(ctx);
112
poly1305_blocks_x86_64(ctx, inp, len, padbit);
113
return;
114
}
115
116
do {
117
const unsigned int bytes = min(len, SZ_4K);
118
119
kernel_fpu_begin();
120
if (static_branch_likely(&poly1305_use_avx512))
121
poly1305_blocks_avx512(ctx, inp, bytes, padbit);
122
else if (static_branch_likely(&poly1305_use_avx2))
123
poly1305_blocks_avx2(ctx, inp, bytes, padbit);
124
else
125
poly1305_blocks_avx(ctx, inp, bytes, padbit);
126
kernel_fpu_end();
127
128
len -= bytes;
129
inp += bytes;
130
} while (len);
131
}
132
EXPORT_SYMBOL_GPL(poly1305_blocks_arch);
133
134
void poly1305_emit_arch(const struct poly1305_state *ctx,
135
u8 mac[POLY1305_DIGEST_SIZE], const u32 nonce[4])
136
{
137
if (!static_branch_likely(&poly1305_use_avx))
138
poly1305_emit_x86_64(ctx, mac, nonce);
139
else
140
poly1305_emit_avx(ctx, mac, nonce);
141
}
142
EXPORT_SYMBOL_GPL(poly1305_emit_arch);
143
144
bool poly1305_is_arch_optimized(void)
145
{
146
return static_key_enabled(&poly1305_use_avx);
147
}
148
EXPORT_SYMBOL(poly1305_is_arch_optimized);
149
150
static int __init poly1305_simd_mod_init(void)
151
{
152
if (boot_cpu_has(X86_FEATURE_AVX) &&
153
cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM, NULL))
154
static_branch_enable(&poly1305_use_avx);
155
if (boot_cpu_has(X86_FEATURE_AVX) && boot_cpu_has(X86_FEATURE_AVX2) &&
156
cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM, NULL))
157
static_branch_enable(&poly1305_use_avx2);
158
if (boot_cpu_has(X86_FEATURE_AVX) && boot_cpu_has(X86_FEATURE_AVX2) &&
159
boot_cpu_has(X86_FEATURE_AVX512F) &&
160
cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM | XFEATURE_MASK_AVX512, NULL) &&
161
/* Skylake downclocks unacceptably much when using zmm, but later generations are fast. */
162
boot_cpu_data.x86_vfm != INTEL_SKYLAKE_X)
163
static_branch_enable(&poly1305_use_avx512);
164
return 0;
165
}
166
subsys_initcall(poly1305_simd_mod_init);
167
168
static void __exit poly1305_simd_mod_exit(void)
169
{
170
}
171
module_exit(poly1305_simd_mod_exit);
172
173
MODULE_LICENSE("GPL");
174
MODULE_AUTHOR("Jason A. Donenfeld <[email protected]>");
175
MODULE_DESCRIPTION("Poly1305 authenticator");
176
177