|
| 1 | +// SPDX-License-Identifier: GPL-2.0-or-later |
| 2 | +/* |
| 3 | + * PowerPC P10 (ppc64le) accelerated ChaCha and XChaCha stream ciphers, |
| 4 | + * including ChaCha20 (RFC7539) |
| 5 | + * |
| 6 | + * Copyright 2023- IBM Corp. All rights reserved. |
| 7 | + */ |
| 8 | + |
| 9 | +#include <crypto/algapi.h> |
| 10 | +#include <crypto/internal/chacha.h> |
| 11 | +#include <crypto/internal/simd.h> |
| 12 | +#include <crypto/internal/skcipher.h> |
| 13 | +#include <linux/kernel.h> |
| 14 | +#include <linux/module.h> |
| 15 | +#include <linux/cpufeature.h> |
| 16 | +#include <linux/sizes.h> |
| 17 | +#include <asm/simd.h> |
| 18 | +#include <asm/switch_to.h> |
| 19 | + |
| 20 | +asmlinkage void chacha_p10le_8x(u32 *state, u8 *dst, const u8 *src, |
| 21 | + unsigned int len, int nrounds); |
| 22 | + |
| 23 | +static __ro_after_init DEFINE_STATIC_KEY_FALSE(have_p10); |
| 24 | + |
| 25 | +static void vsx_begin(void) |
| 26 | +{ |
| 27 | + preempt_disable(); |
| 28 | + enable_kernel_vsx(); |
| 29 | +} |
| 30 | + |
| 31 | +static void vsx_end(void) |
| 32 | +{ |
| 33 | + disable_kernel_vsx(); |
| 34 | + preempt_enable(); |
| 35 | +} |
| 36 | + |
| 37 | +static void chacha_p10_do_8x(u32 *state, u8 *dst, const u8 *src, |
| 38 | + unsigned int bytes, int nrounds) |
| 39 | +{ |
| 40 | + unsigned int l = bytes & ~0x0FF; |
| 41 | + |
| 42 | + if (l > 0) { |
| 43 | + chacha_p10le_8x(state, dst, src, l, nrounds); |
| 44 | + bytes -= l; |
| 45 | + src += l; |
| 46 | + dst += l; |
| 47 | + state[12] += l / CHACHA_BLOCK_SIZE; |
| 48 | + } |
| 49 | + |
| 50 | + if (bytes > 0) |
| 51 | + chacha_crypt_generic(state, dst, src, bytes, nrounds); |
| 52 | +} |
| 53 | + |
| 54 | +void hchacha_block_arch(const u32 *state, u32 *stream, int nrounds) |
| 55 | +{ |
| 56 | + hchacha_block_generic(state, stream, nrounds); |
| 57 | +} |
| 58 | +EXPORT_SYMBOL(hchacha_block_arch); |
| 59 | + |
| 60 | +void chacha_init_arch(u32 *state, const u32 *key, const u8 *iv) |
| 61 | +{ |
| 62 | + chacha_init_generic(state, key, iv); |
| 63 | +} |
| 64 | +EXPORT_SYMBOL(chacha_init_arch); |
| 65 | + |
| 66 | +void chacha_crypt_arch(u32 *state, u8 *dst, const u8 *src, unsigned int bytes, |
| 67 | + int nrounds) |
| 68 | +{ |
| 69 | + if (!static_branch_likely(&have_p10) || bytes <= CHACHA_BLOCK_SIZE || |
| 70 | + !crypto_simd_usable()) |
| 71 | + return chacha_crypt_generic(state, dst, src, bytes, nrounds); |
| 72 | + |
| 73 | + do { |
| 74 | + unsigned int todo = min_t(unsigned int, bytes, SZ_4K); |
| 75 | + |
| 76 | + vsx_begin(); |
| 77 | + chacha_p10_do_8x(state, dst, src, todo, nrounds); |
| 78 | + vsx_end(); |
| 79 | + |
| 80 | + bytes -= todo; |
| 81 | + src += todo; |
| 82 | + dst += todo; |
| 83 | + } while (bytes); |
| 84 | +} |
| 85 | +EXPORT_SYMBOL(chacha_crypt_arch); |
| 86 | + |
| 87 | +static int chacha_p10_stream_xor(struct skcipher_request *req, |
| 88 | + const struct chacha_ctx *ctx, const u8 *iv) |
| 89 | +{ |
| 90 | + struct skcipher_walk walk; |
| 91 | + u32 state[16]; |
| 92 | + int err; |
| 93 | + |
| 94 | + err = skcipher_walk_virt(&walk, req, false); |
| 95 | + if (err) |
| 96 | + return err; |
| 97 | + |
| 98 | + chacha_init_generic(state, ctx->key, iv); |
| 99 | + |
| 100 | + while (walk.nbytes > 0) { |
| 101 | + unsigned int nbytes = walk.nbytes; |
| 102 | + |
| 103 | + if (nbytes < walk.total) |
| 104 | + nbytes = rounddown(nbytes, walk.stride); |
| 105 | + |
| 106 | + if (!crypto_simd_usable()) { |
| 107 | + chacha_crypt_generic(state, walk.dst.virt.addr, |
| 108 | + walk.src.virt.addr, nbytes, |
| 109 | + ctx->nrounds); |
| 110 | + } else { |
| 111 | + vsx_begin(); |
| 112 | + chacha_p10_do_8x(state, walk.dst.virt.addr, |
| 113 | + walk.src.virt.addr, nbytes, ctx->nrounds); |
| 114 | + vsx_end(); |
| 115 | + } |
| 116 | + err = skcipher_walk_done(&walk, walk.nbytes - nbytes); |
| 117 | + if (err) |
| 118 | + break; |
| 119 | + } |
| 120 | + |
| 121 | + return err; |
| 122 | +} |
| 123 | + |
| 124 | +static int chacha_p10(struct skcipher_request *req) |
| 125 | +{ |
| 126 | + struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req); |
| 127 | + struct chacha_ctx *ctx = crypto_skcipher_ctx(tfm); |
| 128 | + |
| 129 | + return chacha_p10_stream_xor(req, ctx, req->iv); |
| 130 | +} |
| 131 | + |
| 132 | +static int xchacha_p10(struct skcipher_request *req) |
| 133 | +{ |
| 134 | + struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req); |
| 135 | + struct chacha_ctx *ctx = crypto_skcipher_ctx(tfm); |
| 136 | + struct chacha_ctx subctx; |
| 137 | + u32 state[16]; |
| 138 | + u8 real_iv[16]; |
| 139 | + |
| 140 | + chacha_init_generic(state, ctx->key, req->iv); |
| 141 | + hchacha_block_arch(state, subctx.key, ctx->nrounds); |
| 142 | + subctx.nrounds = ctx->nrounds; |
| 143 | + |
| 144 | + memcpy(&real_iv[0], req->iv + 24, 8); |
| 145 | + memcpy(&real_iv[8], req->iv + 16, 8); |
| 146 | + return chacha_p10_stream_xor(req, &subctx, real_iv); |
| 147 | +} |
| 148 | + |
| 149 | +static struct skcipher_alg algs[] = { |
| 150 | + { |
| 151 | + .base.cra_name = "chacha20", |
| 152 | + .base.cra_driver_name = "chacha20-p10", |
| 153 | + .base.cra_priority = 300, |
| 154 | + .base.cra_blocksize = 1, |
| 155 | + .base.cra_ctxsize = sizeof(struct chacha_ctx), |
| 156 | + .base.cra_module = THIS_MODULE, |
| 157 | + |
| 158 | + .min_keysize = CHACHA_KEY_SIZE, |
| 159 | + .max_keysize = CHACHA_KEY_SIZE, |
| 160 | + .ivsize = CHACHA_IV_SIZE, |
| 161 | + .chunksize = CHACHA_BLOCK_SIZE, |
| 162 | + .setkey = chacha20_setkey, |
| 163 | + .encrypt = chacha_p10, |
| 164 | + .decrypt = chacha_p10, |
| 165 | + }, { |
| 166 | + .base.cra_name = "xchacha20", |
| 167 | + .base.cra_driver_name = "xchacha20-p10", |
| 168 | + .base.cra_priority = 300, |
| 169 | + .base.cra_blocksize = 1, |
| 170 | + .base.cra_ctxsize = sizeof(struct chacha_ctx), |
| 171 | + .base.cra_module = THIS_MODULE, |
| 172 | + |
| 173 | + .min_keysize = CHACHA_KEY_SIZE, |
| 174 | + .max_keysize = CHACHA_KEY_SIZE, |
| 175 | + .ivsize = XCHACHA_IV_SIZE, |
| 176 | + .chunksize = CHACHA_BLOCK_SIZE, |
| 177 | + .setkey = chacha20_setkey, |
| 178 | + .encrypt = xchacha_p10, |
| 179 | + .decrypt = xchacha_p10, |
| 180 | + }, { |
| 181 | + .base.cra_name = "xchacha12", |
| 182 | + .base.cra_driver_name = "xchacha12-p10", |
| 183 | + .base.cra_priority = 300, |
| 184 | + .base.cra_blocksize = 1, |
| 185 | + .base.cra_ctxsize = sizeof(struct chacha_ctx), |
| 186 | + .base.cra_module = THIS_MODULE, |
| 187 | + |
| 188 | + .min_keysize = CHACHA_KEY_SIZE, |
| 189 | + .max_keysize = CHACHA_KEY_SIZE, |
| 190 | + .ivsize = XCHACHA_IV_SIZE, |
| 191 | + .chunksize = CHACHA_BLOCK_SIZE, |
| 192 | + .setkey = chacha12_setkey, |
| 193 | + .encrypt = xchacha_p10, |
| 194 | + .decrypt = xchacha_p10, |
| 195 | + } |
| 196 | +}; |
| 197 | + |
| 198 | +static int __init chacha_p10_init(void) |
| 199 | +{ |
| 200 | + static_branch_enable(&have_p10); |
| 201 | + |
| 202 | + return crypto_register_skciphers(algs, ARRAY_SIZE(algs)); |
| 203 | +} |
| 204 | + |
| 205 | +static void __exit chacha_p10_exit(void) |
| 206 | +{ |
| 207 | + crypto_unregister_skciphers(algs, ARRAY_SIZE(algs)); |
| 208 | +} |
| 209 | + |
| 210 | +module_cpu_feature_match(PPC_MODULE_FEATURE_P10, chacha_p10_init); |
| 211 | +module_exit(chacha_p10_exit); |
| 212 | + |
| 213 | +MODULE_DESCRIPTION("ChaCha and XChaCha stream ciphers (P10 accelerated)"); |
| 214 | +MODULE_AUTHOR( "Danny Tsen <[email protected]>"); |
| 215 | +MODULE_LICENSE("GPL v2"); |
| 216 | +MODULE_ALIAS_CRYPTO("chacha20"); |
| 217 | +MODULE_ALIAS_CRYPTO("chacha20-p10"); |
| 218 | +MODULE_ALIAS_CRYPTO("xchacha20"); |
| 219 | +MODULE_ALIAS_CRYPTO("xchacha20-p10"); |
| 220 | +MODULE_ALIAS_CRYPTO("xchacha12"); |
| 221 | +MODULE_ALIAS_CRYPTO("xchacha12-p10"); |
0 commit comments