1*4882a593Smuzhiyun // SPDX-License-Identifier: GPL-2.0 OR MIT
2*4882a593Smuzhiyun /*
3*4882a593Smuzhiyun * Copyright (C) 2015-2019 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
4*4882a593Smuzhiyun */
5*4882a593Smuzhiyun
6*4882a593Smuzhiyun #include <crypto/internal/blake2s.h>
7*4882a593Smuzhiyun #include <crypto/internal/simd.h>
8*4882a593Smuzhiyun
9*4882a593Smuzhiyun #include <linux/types.h>
10*4882a593Smuzhiyun #include <linux/jump_label.h>
11*4882a593Smuzhiyun #include <linux/kernel.h>
12*4882a593Smuzhiyun #include <linux/module.h>
13*4882a593Smuzhiyun #include <linux/sizes.h>
14*4882a593Smuzhiyun
15*4882a593Smuzhiyun #include <asm/cpufeature.h>
16*4882a593Smuzhiyun #include <asm/fpu/api.h>
17*4882a593Smuzhiyun #include <asm/processor.h>
18*4882a593Smuzhiyun #include <asm/simd.h>
19*4882a593Smuzhiyun
20*4882a593Smuzhiyun asmlinkage void blake2s_compress_ssse3(struct blake2s_state *state,
21*4882a593Smuzhiyun const u8 *block, const size_t nblocks,
22*4882a593Smuzhiyun const u32 inc);
23*4882a593Smuzhiyun asmlinkage void blake2s_compress_avx512(struct blake2s_state *state,
24*4882a593Smuzhiyun const u8 *block, const size_t nblocks,
25*4882a593Smuzhiyun const u32 inc);
26*4882a593Smuzhiyun
27*4882a593Smuzhiyun static __ro_after_init DEFINE_STATIC_KEY_FALSE(blake2s_use_ssse3);
28*4882a593Smuzhiyun static __ro_after_init DEFINE_STATIC_KEY_FALSE(blake2s_use_avx512);
29*4882a593Smuzhiyun
blake2s_compress(struct blake2s_state * state,const u8 * block,size_t nblocks,const u32 inc)30*4882a593Smuzhiyun void blake2s_compress(struct blake2s_state *state, const u8 *block,
31*4882a593Smuzhiyun size_t nblocks, const u32 inc)
32*4882a593Smuzhiyun {
33*4882a593Smuzhiyun /* SIMD disables preemption, so relax after processing each page. */
34*4882a593Smuzhiyun BUILD_BUG_ON(SZ_4K / BLAKE2S_BLOCK_SIZE < 8);
35*4882a593Smuzhiyun
36*4882a593Smuzhiyun if (!static_branch_likely(&blake2s_use_ssse3) || !crypto_simd_usable()) {
37*4882a593Smuzhiyun blake2s_compress_generic(state, block, nblocks, inc);
38*4882a593Smuzhiyun return;
39*4882a593Smuzhiyun }
40*4882a593Smuzhiyun
41*4882a593Smuzhiyun do {
42*4882a593Smuzhiyun const size_t blocks = min_t(size_t, nblocks,
43*4882a593Smuzhiyun SZ_4K / BLAKE2S_BLOCK_SIZE);
44*4882a593Smuzhiyun
45*4882a593Smuzhiyun kernel_fpu_begin();
46*4882a593Smuzhiyun if (IS_ENABLED(CONFIG_AS_AVX512) &&
47*4882a593Smuzhiyun static_branch_likely(&blake2s_use_avx512))
48*4882a593Smuzhiyun blake2s_compress_avx512(state, block, blocks, inc);
49*4882a593Smuzhiyun else
50*4882a593Smuzhiyun blake2s_compress_ssse3(state, block, blocks, inc);
51*4882a593Smuzhiyun kernel_fpu_end();
52*4882a593Smuzhiyun
53*4882a593Smuzhiyun nblocks -= blocks;
54*4882a593Smuzhiyun block += blocks * BLAKE2S_BLOCK_SIZE;
55*4882a593Smuzhiyun } while (nblocks);
56*4882a593Smuzhiyun }
57*4882a593Smuzhiyun EXPORT_SYMBOL(blake2s_compress);
58*4882a593Smuzhiyun
blake2s_mod_init(void)59*4882a593Smuzhiyun static int __init blake2s_mod_init(void)
60*4882a593Smuzhiyun {
61*4882a593Smuzhiyun if (boot_cpu_has(X86_FEATURE_SSSE3))
62*4882a593Smuzhiyun static_branch_enable(&blake2s_use_ssse3);
63*4882a593Smuzhiyun
64*4882a593Smuzhiyun if (IS_ENABLED(CONFIG_AS_AVX512) &&
65*4882a593Smuzhiyun boot_cpu_has(X86_FEATURE_AVX) &&
66*4882a593Smuzhiyun boot_cpu_has(X86_FEATURE_AVX2) &&
67*4882a593Smuzhiyun boot_cpu_has(X86_FEATURE_AVX512F) &&
68*4882a593Smuzhiyun boot_cpu_has(X86_FEATURE_AVX512VL) &&
69*4882a593Smuzhiyun cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM |
70*4882a593Smuzhiyun XFEATURE_MASK_AVX512, NULL))
71*4882a593Smuzhiyun static_branch_enable(&blake2s_use_avx512);
72*4882a593Smuzhiyun
73*4882a593Smuzhiyun return 0;
74*4882a593Smuzhiyun }
75*4882a593Smuzhiyun
76*4882a593Smuzhiyun module_init(blake2s_mod_init);
77*4882a593Smuzhiyun
78*4882a593Smuzhiyun MODULE_LICENSE("GPL v2");
79