1*39ee3970SEric Biggers /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2*39ee3970SEric Biggers /*
3*39ee3970SEric Biggers * Copyright (C) 2015-2019 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
4*39ee3970SEric Biggers */
5*39ee3970SEric Biggers
6*39ee3970SEric Biggers #include <asm/cpufeature.h>
7*39ee3970SEric Biggers #include <asm/fpu/api.h>
8*39ee3970SEric Biggers #include <asm/processor.h>
9*39ee3970SEric Biggers #include <asm/simd.h>
10*39ee3970SEric Biggers #include <linux/jump_label.h>
11*39ee3970SEric Biggers #include <linux/kernel.h>
12*39ee3970SEric Biggers #include <linux/sizes.h>
13*39ee3970SEric Biggers
14*39ee3970SEric Biggers asmlinkage void blake2s_compress_ssse3(struct blake2s_state *state,
15*39ee3970SEric Biggers const u8 *block, const size_t nblocks,
16*39ee3970SEric Biggers const u32 inc);
17*39ee3970SEric Biggers asmlinkage void blake2s_compress_avx512(struct blake2s_state *state,
18*39ee3970SEric Biggers const u8 *block, const size_t nblocks,
19*39ee3970SEric Biggers const u32 inc);
20*39ee3970SEric Biggers
21*39ee3970SEric Biggers static __ro_after_init DEFINE_STATIC_KEY_FALSE(blake2s_use_ssse3);
22*39ee3970SEric Biggers static __ro_after_init DEFINE_STATIC_KEY_FALSE(blake2s_use_avx512);
23*39ee3970SEric Biggers
blake2s_compress(struct blake2s_state * state,const u8 * block,size_t nblocks,const u32 inc)24*39ee3970SEric Biggers static void blake2s_compress(struct blake2s_state *state, const u8 *block,
25*39ee3970SEric Biggers size_t nblocks, const u32 inc)
26*39ee3970SEric Biggers {
27*39ee3970SEric Biggers /* SIMD disables preemption, so relax after processing each page. */
28*39ee3970SEric Biggers BUILD_BUG_ON(SZ_4K / BLAKE2S_BLOCK_SIZE < 8);
29*39ee3970SEric Biggers
30*39ee3970SEric Biggers if (!static_branch_likely(&blake2s_use_ssse3) || !may_use_simd()) {
31*39ee3970SEric Biggers blake2s_compress_generic(state, block, nblocks, inc);
32*39ee3970SEric Biggers return;
33*39ee3970SEric Biggers }
34*39ee3970SEric Biggers
35*39ee3970SEric Biggers do {
36*39ee3970SEric Biggers const size_t blocks = min_t(size_t, nblocks,
37*39ee3970SEric Biggers SZ_4K / BLAKE2S_BLOCK_SIZE);
38*39ee3970SEric Biggers
39*39ee3970SEric Biggers kernel_fpu_begin();
40*39ee3970SEric Biggers if (static_branch_likely(&blake2s_use_avx512))
41*39ee3970SEric Biggers blake2s_compress_avx512(state, block, blocks, inc);
42*39ee3970SEric Biggers else
43*39ee3970SEric Biggers blake2s_compress_ssse3(state, block, blocks, inc);
44*39ee3970SEric Biggers kernel_fpu_end();
45*39ee3970SEric Biggers
46*39ee3970SEric Biggers nblocks -= blocks;
47*39ee3970SEric Biggers block += blocks * BLAKE2S_BLOCK_SIZE;
48*39ee3970SEric Biggers } while (nblocks);
49*39ee3970SEric Biggers }
50*39ee3970SEric Biggers
51*39ee3970SEric Biggers #define blake2s_mod_init_arch blake2s_mod_init_arch
blake2s_mod_init_arch(void)52*39ee3970SEric Biggers static void blake2s_mod_init_arch(void)
53*39ee3970SEric Biggers {
54*39ee3970SEric Biggers if (boot_cpu_has(X86_FEATURE_SSSE3))
55*39ee3970SEric Biggers static_branch_enable(&blake2s_use_ssse3);
56*39ee3970SEric Biggers
57*39ee3970SEric Biggers if (boot_cpu_has(X86_FEATURE_AVX) &&
58*39ee3970SEric Biggers boot_cpu_has(X86_FEATURE_AVX2) &&
59*39ee3970SEric Biggers boot_cpu_has(X86_FEATURE_AVX512F) &&
60*39ee3970SEric Biggers boot_cpu_has(X86_FEATURE_AVX512VL) &&
61*39ee3970SEric Biggers cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM |
62*39ee3970SEric Biggers XFEATURE_MASK_AVX512, NULL))
63*39ee3970SEric Biggers static_branch_enable(&blake2s_use_avx512);
64*39ee3970SEric Biggers }
65