arch/arm/crypto/chacha-glue.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * ARM NEON accelerated ChaCha and XChaCha stream ciphers,
   4  * including ChaCha20 (RFC7539)
   5  *
   6  * Copyright (C) 2016-2019 Linaro, Ltd. <ard.biesheuvel@linaro.org>
   7  * Copyright (C) 2015 Martin Willi
   8  */
   9
  10 #include <crypto/algapi.h>
  11 #include <crypto/internal/chacha.h>
  12 #include <crypto/internal/simd.h>
  13 #include <crypto/internal/skcipher.h>
  14 #include <linux/jump_label.h>
  15 #include <linux/kernel.h>
  16 #include <linux/module.h>
  17
  18 #include <asm/cputype.h>
  19 #include <asm/hwcap.h>
  20 #include <asm/neon.h>
  21 #include <asm/simd.h>
  22
  23 asmlinkage void chacha_block_xor_neon(const u32 *state, u8 *dst, const u8 *src,
  24                                       int nrounds);
  25 asmlinkage void chacha_4block_xor_neon(const u32 *state, u8 *dst, const u8 *src,
  26                                        int nrounds, unsigned int nbytes);
  27 asmlinkage void hchacha_block_arm(const u32 *state, u32 *out, int nrounds);
  28 asmlinkage void hchacha_block_neon(const u32 *state, u32 *out, int nrounds);
  29
  30 asmlinkage void chacha_doarm(u8 *dst, const u8 *src, unsigned int bytes,
  31                              const u32 *state, int nrounds);
  32
  33 static __ro_after_init DEFINE_STATIC_KEY_FALSE(use_neon);
  34
  35 static inline bool neon_usable(void)
  36 {
  37         return static_branch_likely(&use_neon) && crypto_simd_usable();
  38 }
  39
  40 static void chacha_doneon(u32 *state, u8 *dst, const u8 *src,
  41                           unsigned int bytes, int nrounds)
  42 {
  43         u8 buf[CHACHA_BLOCK_SIZE];
  44
  45         while (bytes > CHACHA_BLOCK_SIZE) {
  46                 unsigned int l = min(bytes, CHACHA_BLOCK_SIZE * 4U);
  47
  48                 chacha_4block_xor_neon(state, dst, src, nrounds, l);
  49                 bytes -= l;
  50                 src += l;
  51                 dst += l;
  52                 state[12] += DIV_ROUND_UP(l, CHACHA_BLOCK_SIZE);
  53         }
  54         if (bytes) {
  55                 const u8 *s = src;
  56                 u8 *d = dst;
  57
  58                 if (bytes != CHACHA_BLOCK_SIZE)
  59                         s = d = memcpy(buf, src, bytes);
  60                 chacha_block_xor_neon(state, d, s, nrounds);
  61                 if (d != dst)
  62                         memcpy(dst, buf, bytes);
  63         }
  64 }
  65
  66 void hchacha_block_arch(const u32 *state, u32 *stream, int nrounds)
  67 {
  68         if (!IS_ENABLED(CONFIG_KERNEL_MODE_NEON) || !neon_usable()) {
  69                 hchacha_block_arm(state, stream, nrounds);
  70         } else {
  71                 kernel_neon_begin();
  72                 hchacha_block_neon(state, stream, nrounds);
  73                 kernel_neon_end();
  74         }
  75 }
  76 EXPORT_SYMBOL(hchacha_block_arch);
  77
  78 void chacha_init_arch(u32 *state, const u32 *key, const u8 *iv)
  79 {
  80         chacha_init_generic(state, key, iv);
  81 }
  82 EXPORT_SYMBOL(chacha_init_arch);
  83
  84 void chacha_crypt_arch(u32 *state, u8 *dst, const u8 *src, unsigned int bytes,
  85                        int nrounds)
  86 {
  87         if (!IS_ENABLED(CONFIG_KERNEL_MODE_NEON) || !neon_usable() ||
  88             bytes <= CHACHA_BLOCK_SIZE) {
  89                 chacha_doarm(dst, src, bytes, state, nrounds);
  90                 state[12] += DIV_ROUND_UP(bytes, CHACHA_BLOCK_SIZE);
  91                 return;
  92         }
  93
  94         do {
  95                 unsigned int todo = min_t(unsigned int, bytes, SZ_4K);
  96
  97                 kernel_neon_begin();
  98                 chacha_doneon(state, dst, src, todo, nrounds);
  99                 kernel_neon_end();
 100
 101                 bytes -= todo;
 102                 src += todo;
 103                 dst += todo;
 104         } while (bytes);
 105 }
 106 EXPORT_SYMBOL(chacha_crypt_arch);
 107
 108 static int chacha_stream_xor(struct skcipher_request *req,
 109                              const struct chacha_ctx *ctx, const u8 *iv,
 110                              bool neon)
 111 {
 112         struct skcipher_walk walk;
 113         u32 state[16];
 114         int err;
 115
 116         err = skcipher_walk_virt(&walk, req, false);
 117
 118         chacha_init_generic(state, ctx->key, iv);
 119
 120         while (walk.nbytes > 0) {
 121                 unsigned int nbytes = walk.nbytes;
 122
 123                 if (nbytes < walk.total)
 124                         nbytes = round_down(nbytes, walk.stride);
 125
 126                 if (!IS_ENABLED(CONFIG_KERNEL_MODE_NEON) || !neon) {
 127                         chacha_doarm(walk.dst.virt.addr, walk.src.virt.addr,
 128                                      nbytes, state, ctx->nrounds);
 129                         state[12] += DIV_ROUND_UP(nbytes, CHACHA_BLOCK_SIZE);
 130                 } else {
 131                         kernel_neon_begin();
 132                         chacha_doneon(state, walk.dst.virt.addr,
 133                                       walk.src.virt.addr, nbytes, ctx->nrounds);
 134                         kernel_neon_end();
 135                 }
 136                 err = skcipher_walk_done(&walk, walk.nbytes - nbytes);
 137         }
 138
 139         return err;
 140 }
 141
 142 static int do_chacha(struct skcipher_request *req, bool neon)
 143 {
 144         struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 145         struct chacha_ctx *ctx = crypto_skcipher_ctx(tfm);
 146
 147         return chacha_stream_xor(req, ctx, req->iv, neon);
 148 }
 149
 150 static int chacha_arm(struct skcipher_request *req)
 151 {
 152         return do_chacha(req, false);
 153 }
 154
 155 static int chacha_neon(struct skcipher_request *req)
 156 {
 157         return do_chacha(req, neon_usable());
 158 }
 159
 160 static int do_xchacha(struct skcipher_request *req, bool neon)
 161 {
 162         struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
 163         struct chacha_ctx *ctx = crypto_skcipher_ctx(tfm);
 164         struct chacha_ctx subctx;
 165         u32 state[16];
 166         u8 real_iv[16];
 167
 168         chacha_init_generic(state, ctx->key, req->iv);
 169
 170         if (!IS_ENABLED(CONFIG_KERNEL_MODE_NEON) || !neon) {
 171                 hchacha_block_arm(state, subctx.key, ctx->nrounds);
 172         } else {
 173                 kernel_neon_begin();
 174                 hchacha_block_neon(state, subctx.key, ctx->nrounds);
 175                 kernel_neon_end();
 176         }
 177         subctx.nrounds = ctx->nrounds;
 178
 179         memcpy(&real_iv[0], req->iv + 24, 8);
 180         memcpy(&real_iv[8], req->iv + 16, 8);
 181         return chacha_stream_xor(req, &subctx, real_iv, neon);
 182 }
 183
 184 static int xchacha_arm(struct skcipher_request *req)
 185 {
 186         return do_xchacha(req, false);
 187 }
 188
 189 static int xchacha_neon(struct skcipher_request *req)
 190 {
 191         return do_xchacha(req, neon_usable());
 192 }
 193
 194 static struct skcipher_alg arm_algs[] = {
 195         {
 196                 .base.cra_name          = "chacha20",
 197                 .base.cra_driver_name   = "chacha20-arm",
 198                 .base.cra_priority      = 200,
 199                 .base.cra_blocksize     = 1,
 200                 .base.cra_ctxsize       = sizeof(struct chacha_ctx),
 201                 .base.cra_module        = THIS_MODULE,
 202
 203                 .min_keysize            = CHACHA_KEY_SIZE,
 204                 .max_keysize            = CHACHA_KEY_SIZE,
 205                 .ivsize                 = CHACHA_IV_SIZE,
 206                 .chunksize              = CHACHA_BLOCK_SIZE,
 207                 .setkey                 = chacha20_setkey,
 208                 .encrypt                = chacha_arm,
 209                 .decrypt                = chacha_arm,
 210         }, {
 211                 .base.cra_name          = "xchacha20",
 212                 .base.cra_driver_name   = "xchacha20-arm",
 213                 .base.cra_priority      = 200,
 214                 .base.cra_blocksize     = 1,
 215                 .base.cra_ctxsize       = sizeof(struct chacha_ctx),
 216                 .base.cra_module        = THIS_MODULE,
 217
 218                 .min_keysize            = CHACHA_KEY_SIZE,
 219                 .max_keysize            = CHACHA_KEY_SIZE,
 220                 .ivsize                 = XCHACHA_IV_SIZE,
 221                 .chunksize              = CHACHA_BLOCK_SIZE,
 222                 .setkey                 = chacha20_setkey,
 223                 .encrypt                = xchacha_arm,
 224                 .decrypt                = xchacha_arm,
 225         }, {
 226                 .base.cra_name          = "xchacha12",
 227                 .base.cra_driver_name   = "xchacha12-arm",
 228                 .base.cra_priority      = 200,
 229                 .base.cra_blocksize     = 1,
 230                 .base.cra_ctxsize       = sizeof(struct chacha_ctx),
 231                 .base.cra_module        = THIS_MODULE,
 232
 233                 .min_keysize            = CHACHA_KEY_SIZE,
 234                 .max_keysize            = CHACHA_KEY_SIZE,
 235                 .ivsize                 = XCHACHA_IV_SIZE,
 236                 .chunksize              = CHACHA_BLOCK_SIZE,
 237                 .setkey                 = chacha12_setkey,
 238                 .encrypt                = xchacha_arm,
 239                 .decrypt                = xchacha_arm,
 240         },
 241 };
 242
 243 static struct skcipher_alg neon_algs[] = {
 244         {
 245                 .base.cra_name          = "chacha20",
 246                 .base.cra_driver_name   = "chacha20-neon",
 247                 .base.cra_priority      = 300,
 248                 .base.cra_blocksize     = 1,
 249                 .base.cra_ctxsize       = sizeof(struct chacha_ctx),
 250                 .base.cra_module        = THIS_MODULE,
 251
 252                 .min_keysize            = CHACHA_KEY_SIZE,
 253                 .max_keysize            = CHACHA_KEY_SIZE,
 254                 .ivsize                 = CHACHA_IV_SIZE,
 255                 .chunksize              = CHACHA_BLOCK_SIZE,
 256                 .walksize               = 4 * CHACHA_BLOCK_SIZE,
 257                 .setkey                 = chacha20_setkey,
 258                 .encrypt                = chacha_neon,
 259                 .decrypt                = chacha_neon,
 260         }, {
 261                 .base.cra_name          = "xchacha20",
 262                 .base.cra_driver_name   = "xchacha20-neon",
 263                 .base.cra_priority      = 300,
 264                 .base.cra_blocksize     = 1,
 265                 .base.cra_ctxsize       = sizeof(struct chacha_ctx),
 266                 .base.cra_module        = THIS_MODULE,
 267
 268                 .min_keysize            = CHACHA_KEY_SIZE,
 269                 .max_keysize            = CHACHA_KEY_SIZE,
 270                 .ivsize                 = XCHACHA_IV_SIZE,
 271                 .chunksize              = CHACHA_BLOCK_SIZE,
 272                 .walksize               = 4 * CHACHA_BLOCK_SIZE,
 273                 .setkey                 = chacha20_setkey,
 274                 .encrypt                = xchacha_neon,
 275                 .decrypt                = xchacha_neon,
 276         }, {
 277                 .base.cra_name          = "xchacha12",
 278                 .base.cra_driver_name   = "xchacha12-neon",
 279                 .base.cra_priority      = 300,
 280                 .base.cra_blocksize     = 1,
 281                 .base.cra_ctxsize       = sizeof(struct chacha_ctx),
 282                 .base.cra_module        = THIS_MODULE,
 283
 284                 .min_keysize            = CHACHA_KEY_SIZE,
 285                 .max_keysize            = CHACHA_KEY_SIZE,
 286                 .ivsize                 = XCHACHA_IV_SIZE,
 287                 .chunksize              = CHACHA_BLOCK_SIZE,
 288                 .walksize               = 4 * CHACHA_BLOCK_SIZE,
 289                 .setkey                 = chacha12_setkey,
 290                 .encrypt                = xchacha_neon,
 291                 .decrypt                = xchacha_neon,
 292         }
 293 };
 294
 295 static int __init chacha_simd_mod_init(void)
 296 {
 297         int err = 0;
 298
 299         if (IS_REACHABLE(CONFIG_CRYPTO_SKCIPHER)) {
 300                 err = crypto_register_skciphers(arm_algs, ARRAY_SIZE(arm_algs));
 301                 if (err)
 302                         return err;
 303         }
 304
 305         if (IS_ENABLED(CONFIG_KERNEL_MODE_NEON) && (elf_hwcap & HWCAP_NEON)) {
 306                 int i;
 307
 308                 switch (read_cpuid_part()) {
 309                 case ARM_CPU_PART_CORTEX_A7:
 310                 case ARM_CPU_PART_CORTEX_A5:
 311                         /*
 312                          * The Cortex-A7 and Cortex-A5 do not perform well with
 313                          * the NEON implementation but do incredibly with the
 314                          * scalar one and use less power.
 315                          */
 316                         for (i = 0; i < ARRAY_SIZE(neon_algs); i++)
 317                                 neon_algs[i].base.cra_priority = 0;
 318                         break;
 319                 default:
 320                         static_branch_enable(&use_neon);
 321                 }
 322
 323                 if (IS_REACHABLE(CONFIG_CRYPTO_SKCIPHER)) {
 324                         err = crypto_register_skciphers(neon_algs, ARRAY_SIZE(neon_algs));
 325                         if (err)
 326                                 crypto_unregister_skciphers(arm_algs, ARRAY_SIZE(arm_algs));
 327                 }
 328         }
 329         return err;
 330 }
 331
 332 static void __exit chacha_simd_mod_fini(void)
 333 {
 334         if (IS_REACHABLE(CONFIG_CRYPTO_SKCIPHER)) {
 335                 crypto_unregister_skciphers(arm_algs, ARRAY_SIZE(arm_algs));
 336                 if (IS_ENABLED(CONFIG_KERNEL_MODE_NEON) && (elf_hwcap & HWCAP_NEON))
 337                         crypto_unregister_skciphers(neon_algs, ARRAY_SIZE(neon_algs));
 338         }
 339 }
 340
 341 module_init(chacha_simd_mod_init);
 342 module_exit(chacha_simd_mod_fini);
 343
 344 MODULE_DESCRIPTION("ChaCha and XChaCha stream ciphers (scalar and NEON accelerated)");
 345 MODULE_AUTHOR("Ard Biesheuvel <ard.biesheuvel@linaro.org>");
 346 MODULE_LICENSE("GPL v2");
 347 MODULE_ALIAS_CRYPTO("chacha20");
 348 MODULE_ALIAS_CRYPTO("chacha20-arm");
 349 MODULE_ALIAS_CRYPTO("xchacha20");
 350 MODULE_ALIAS_CRYPTO("xchacha20-arm");
 351 MODULE_ALIAS_CRYPTO("xchacha12");
 352 MODULE_ALIAS_CRYPTO("xchacha12-arm");
 353 #ifdef CONFIG_KERNEL_MODE_NEON
 354 MODULE_ALIAS_CRYPTO("chacha20-neon");
 355 MODULE_ALIAS_CRYPTO("xchacha20-neon");
 356 MODULE_ALIAS_CRYPTO("xchacha12-neon");
 357 #endif