1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* 3 * Copyright (C) 2015-2019 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved. 4 */ 5 6 #include <crypto/algapi.h> 7 #include <crypto/internal/hash.h> 8 #include <crypto/internal/poly1305.h> 9 #include <crypto/internal/simd.h> 10 #include <linux/crypto.h> 11 #include <linux/jump_label.h> 12 #include <linux/kernel.h> 13 #include <linux/module.h> 14 #include <asm/intel-family.h> 15 #include <asm/simd.h> 16 17 asmlinkage void poly1305_init_x86_64(void *ctx, 18 const u8 key[POLY1305_KEY_SIZE]); 19 asmlinkage void poly1305_blocks_x86_64(void *ctx, const u8 *inp, 20 const size_t len, const u32 padbit); 21 asmlinkage void poly1305_emit_x86_64(void *ctx, u8 mac[POLY1305_DIGEST_SIZE], 22 const u32 nonce[4]); 23 asmlinkage void poly1305_emit_avx(void *ctx, u8 mac[POLY1305_DIGEST_SIZE], 24 const u32 nonce[4]); 25 asmlinkage void poly1305_blocks_avx(void *ctx, const u8 *inp, const size_t len, 26 const u32 padbit); 27 asmlinkage void poly1305_blocks_avx2(void *ctx, const u8 *inp, const size_t len, 28 const u32 padbit); 29 asmlinkage void poly1305_blocks_avx512(void *ctx, const u8 *inp, 30 const size_t len, const u32 padbit); 31 32 static __ro_after_init DEFINE_STATIC_KEY_FALSE(poly1305_use_avx); 33 static __ro_after_init DEFINE_STATIC_KEY_FALSE(poly1305_use_avx2); 34 static __ro_after_init DEFINE_STATIC_KEY_FALSE(poly1305_use_avx512); 35 36 struct poly1305_arch_internal { 37 union { 38 struct { 39 u32 h[5]; 40 u32 is_base2_26; 41 }; 42 u64 hs[3]; 43 }; 44 u64 r[2]; 45 u64 pad; 46 struct { u32 r2, r1, r4, r3; } rn[9]; 47 }; 48 49 /* The AVX code uses base 2^26, while the scalar code uses base 2^64. If we hit 50 * the unfortunate situation of using AVX and then having to go back to scalar 51 * -- because the user is silly and has called the update function from two 52 * separate contexts -- then we need to convert back to the original base before 53 * proceeding. It is possible to reason that the initial reduction below is 54 * sufficient given the implementation invariants. However, for an avoidance of 55 * doubt and because this is not performance critical, we do the full reduction 56 * anyway. Z3 proof of below function: https://xn--4db.cc/ltPtHCKN/py 57 */ 58 static void convert_to_base2_64(void *ctx) 59 { 60 struct poly1305_arch_internal *state = ctx; 61 u32 cy; 62 63 if (!state->is_base2_26) 64 return; 65 66 cy = state->h[0] >> 26; state->h[0] &= 0x3ffffff; state->h[1] += cy; 67 cy = state->h[1] >> 26; state->h[1] &= 0x3ffffff; state->h[2] += cy; 68 cy = state->h[2] >> 26; state->h[2] &= 0x3ffffff; state->h[3] += cy; 69 cy = state->h[3] >> 26; state->h[3] &= 0x3ffffff; state->h[4] += cy; 70 state->hs[0] = ((u64)state->h[2] << 52) | ((u64)state->h[1] << 26) | state->h[0]; 71 state->hs[1] = ((u64)state->h[4] << 40) | ((u64)state->h[3] << 14) | (state->h[2] >> 12); 72 state->hs[2] = state->h[4] >> 24; 73 #define ULT(a, b) ((a ^ ((a ^ b) | ((a - b) ^ b))) >> (sizeof(a) * 8 - 1)) 74 cy = (state->hs[2] >> 2) + (state->hs[2] & ~3ULL); 75 state->hs[2] &= 3; 76 state->hs[0] += cy; 77 state->hs[1] += (cy = ULT(state->hs[0], cy)); 78 state->hs[2] += ULT(state->hs[1], cy); 79 #undef ULT 80 state->is_base2_26 = 0; 81 } 82 83 static void poly1305_simd_init(void *ctx, const u8 key[POLY1305_KEY_SIZE]) 84 { 85 poly1305_init_x86_64(ctx, key); 86 } 87 88 static void poly1305_simd_blocks(void *ctx, const u8 *inp, size_t len, 89 const u32 padbit) 90 { 91 struct poly1305_arch_internal *state = ctx; 92 93 /* SIMD disables preemption, so relax after processing each page. */ 94 BUILD_BUG_ON(SZ_4K < POLY1305_BLOCK_SIZE || 95 SZ_4K % POLY1305_BLOCK_SIZE); 96 97 if (!static_branch_likely(&poly1305_use_avx) || 98 (len < (POLY1305_BLOCK_SIZE * 18) && !state->is_base2_26) || 99 !crypto_simd_usable()) { 100 convert_to_base2_64(ctx); 101 poly1305_blocks_x86_64(ctx, inp, len, padbit); 102 return; 103 } 104 105 do { 106 const size_t bytes = min_t(size_t, len, SZ_4K); 107 108 kernel_fpu_begin(); 109 if (IS_ENABLED(CONFIG_AS_AVX512) && static_branch_likely(&poly1305_use_avx512)) 110 poly1305_blocks_avx512(ctx, inp, bytes, padbit); 111 else if (static_branch_likely(&poly1305_use_avx2)) 112 poly1305_blocks_avx2(ctx, inp, bytes, padbit); 113 else 114 poly1305_blocks_avx(ctx, inp, bytes, padbit); 115 kernel_fpu_end(); 116 117 len -= bytes; 118 inp += bytes; 119 } while (len); 120 } 121 122 static void poly1305_simd_emit(void *ctx, u8 mac[POLY1305_DIGEST_SIZE], 123 const u32 nonce[4]) 124 { 125 if (!static_branch_likely(&poly1305_use_avx)) 126 poly1305_emit_x86_64(ctx, mac, nonce); 127 else 128 poly1305_emit_avx(ctx, mac, nonce); 129 } 130 131 void poly1305_init_arch(struct poly1305_desc_ctx *dctx, const u8 *key) 132 { 133 poly1305_simd_init(&dctx->h, key); 134 dctx->s[0] = get_unaligned_le32(&key[16]); 135 dctx->s[1] = get_unaligned_le32(&key[20]); 136 dctx->s[2] = get_unaligned_le32(&key[24]); 137 dctx->s[3] = get_unaligned_le32(&key[28]); 138 dctx->buflen = 0; 139 dctx->sset = true; 140 } 141 EXPORT_SYMBOL(poly1305_init_arch); 142 143 static unsigned int crypto_poly1305_setdctxkey(struct poly1305_desc_ctx *dctx, 144 const u8 *inp, unsigned int len) 145 { 146 unsigned int acc = 0; 147 if (unlikely(!dctx->sset)) { 148 if (!dctx->rset && len >= POLY1305_BLOCK_SIZE) { 149 poly1305_simd_init(&dctx->h, inp); 150 inp += POLY1305_BLOCK_SIZE; 151 len -= POLY1305_BLOCK_SIZE; 152 acc += POLY1305_BLOCK_SIZE; 153 dctx->rset = 1; 154 } 155 if (len >= POLY1305_BLOCK_SIZE) { 156 dctx->s[0] = get_unaligned_le32(&inp[0]); 157 dctx->s[1] = get_unaligned_le32(&inp[4]); 158 dctx->s[2] = get_unaligned_le32(&inp[8]); 159 dctx->s[3] = get_unaligned_le32(&inp[12]); 160 inp += POLY1305_BLOCK_SIZE; 161 len -= POLY1305_BLOCK_SIZE; 162 acc += POLY1305_BLOCK_SIZE; 163 dctx->sset = true; 164 } 165 } 166 return acc; 167 } 168 169 void poly1305_update_arch(struct poly1305_desc_ctx *dctx, const u8 *src, 170 unsigned int srclen) 171 { 172 unsigned int bytes, used; 173 174 if (unlikely(dctx->buflen)) { 175 bytes = min(srclen, POLY1305_BLOCK_SIZE - dctx->buflen); 176 memcpy(dctx->buf + dctx->buflen, src, bytes); 177 src += bytes; 178 srclen -= bytes; 179 dctx->buflen += bytes; 180 181 if (dctx->buflen == POLY1305_BLOCK_SIZE) { 182 if (likely(!crypto_poly1305_setdctxkey(dctx, dctx->buf, POLY1305_BLOCK_SIZE))) 183 poly1305_simd_blocks(&dctx->h, dctx->buf, POLY1305_BLOCK_SIZE, 1); 184 dctx->buflen = 0; 185 } 186 } 187 188 if (likely(srclen >= POLY1305_BLOCK_SIZE)) { 189 bytes = round_down(srclen, POLY1305_BLOCK_SIZE); 190 srclen -= bytes; 191 used = crypto_poly1305_setdctxkey(dctx, src, bytes); 192 if (likely(bytes - used)) 193 poly1305_simd_blocks(&dctx->h, src + used, bytes - used, 1); 194 src += bytes; 195 } 196 197 if (unlikely(srclen)) { 198 dctx->buflen = srclen; 199 memcpy(dctx->buf, src, srclen); 200 } 201 } 202 EXPORT_SYMBOL(poly1305_update_arch); 203 204 void poly1305_final_arch(struct poly1305_desc_ctx *dctx, u8 *dst) 205 { 206 if (unlikely(dctx->buflen)) { 207 dctx->buf[dctx->buflen++] = 1; 208 memset(dctx->buf + dctx->buflen, 0, 209 POLY1305_BLOCK_SIZE - dctx->buflen); 210 poly1305_simd_blocks(&dctx->h, dctx->buf, POLY1305_BLOCK_SIZE, 0); 211 } 212 213 poly1305_simd_emit(&dctx->h, dst, dctx->s); 214 *dctx = (struct poly1305_desc_ctx){}; 215 } 216 EXPORT_SYMBOL(poly1305_final_arch); 217 218 static int crypto_poly1305_init(struct shash_desc *desc) 219 { 220 struct poly1305_desc_ctx *dctx = shash_desc_ctx(desc); 221 222 *dctx = (struct poly1305_desc_ctx){}; 223 return 0; 224 } 225 226 static int crypto_poly1305_update(struct shash_desc *desc, 227 const u8 *src, unsigned int srclen) 228 { 229 struct poly1305_desc_ctx *dctx = shash_desc_ctx(desc); 230 231 poly1305_update_arch(dctx, src, srclen); 232 return 0; 233 } 234 235 static int crypto_poly1305_final(struct shash_desc *desc, u8 *dst) 236 { 237 struct poly1305_desc_ctx *dctx = shash_desc_ctx(desc); 238 239 if (unlikely(!dctx->sset)) 240 return -ENOKEY; 241 242 poly1305_final_arch(dctx, dst); 243 return 0; 244 } 245 246 static struct shash_alg alg = { 247 .digestsize = POLY1305_DIGEST_SIZE, 248 .init = crypto_poly1305_init, 249 .update = crypto_poly1305_update, 250 .final = crypto_poly1305_final, 251 .descsize = sizeof(struct poly1305_desc_ctx), 252 .base = { 253 .cra_name = "poly1305", 254 .cra_driver_name = "poly1305-simd", 255 .cra_priority = 300, 256 .cra_blocksize = POLY1305_BLOCK_SIZE, 257 .cra_module = THIS_MODULE, 258 }, 259 }; 260 261 static int __init poly1305_simd_mod_init(void) 262 { 263 if (boot_cpu_has(X86_FEATURE_AVX) && 264 cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM, NULL)) 265 static_branch_enable(&poly1305_use_avx); 266 if (boot_cpu_has(X86_FEATURE_AVX) && boot_cpu_has(X86_FEATURE_AVX2) && 267 cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM, NULL)) 268 static_branch_enable(&poly1305_use_avx2); 269 if (IS_ENABLED(CONFIG_AS_AVX512) && boot_cpu_has(X86_FEATURE_AVX) && 270 boot_cpu_has(X86_FEATURE_AVX2) && boot_cpu_has(X86_FEATURE_AVX512F) && 271 cpu_has_xfeatures(XFEATURE_MASK_SSE | XFEATURE_MASK_YMM | XFEATURE_MASK_AVX512, NULL) && 272 /* Skylake downclocks unacceptably much when using zmm, but later generations are fast. */ 273 boot_cpu_data.x86_model != INTEL_FAM6_SKYLAKE_X) 274 static_branch_enable(&poly1305_use_avx512); 275 return IS_REACHABLE(CONFIG_CRYPTO_HASH) ? crypto_register_shash(&alg) : 0; 276 } 277 278 static void __exit poly1305_simd_mod_exit(void) 279 { 280 if (IS_REACHABLE(CONFIG_CRYPTO_HASH)) 281 crypto_unregister_shash(&alg); 282 } 283 284 module_init(poly1305_simd_mod_init); 285 module_exit(poly1305_simd_mod_exit); 286 287 MODULE_LICENSE("GPL"); 288 MODULE_AUTHOR("Jason A. Donenfeld <Jason@zx2c4.com>"); 289 MODULE_DESCRIPTION("Poly1305 authenticator"); 290 MODULE_ALIAS_CRYPTO("poly1305"); 291 MODULE_ALIAS_CRYPTO("poly1305-simd"); 292