arch/x86/crypto/crc32c-intel_glue.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Using hardware provided CRC32 instruction to accelerate the CRC32 disposal.
   4  * CRC32C polynomial:0x1EDC6F41(BE)/0x82F63B78(LE)
   5  * CRC32 is a new instruction in Intel SSE4.2, the reference can be found at:
   6  * http://www.intel.com/products/processor/manuals/
   7  * Intel(R) 64 and IA-32 Architectures Software Developer's Manual
   8  * Volume 2A: Instruction Set Reference, A-M
   9  *
  10  * Copyright (C) 2008 Intel Corporation
  11  * Authors: Austin Zhang <austin_zhang@linux.intel.com>
  12  *          Kent Liu <kent.liu@intel.com>
  13  */
  14 #include <linux/init.h>
  15 #include <linux/module.h>
  16 #include <linux/string.h>
  17 #include <linux/kernel.h>
  18 #include <crypto/internal/hash.h>
  19 #include <crypto/internal/simd.h>
  20
  21 #include <asm/cpufeatures.h>
  22 #include <asm/cpu_device_id.h>
  23 #include <asm/simd.h>
  24
  25 #define CHKSUM_BLOCK_SIZE       1
  26 #define CHKSUM_DIGEST_SIZE      4
  27
  28 #define SCALE_F sizeof(unsigned long)
  29
  30 #ifdef CONFIG_X86_64
  31 #define REX_PRE "0x48, "
  32 #else
  33 #define REX_PRE
  34 #endif
  35
  36 #ifdef CONFIG_X86_64
  37 /*
  38  * use carryless multiply version of crc32c when buffer
  39  * size is >= 512 to account
  40  * for fpu state save/restore overhead.
  41  */
  42 #define CRC32C_PCL_BREAKEVEN    512
  43
  44 asmlinkage unsigned int crc_pcl(const u8 *buffer, int len,
  45                                 unsigned int crc_init);
  46 #endif /* CONFIG_X86_64 */
  47
  48 static u32 crc32c_intel_le_hw_byte(u32 crc, unsigned char const *data, size_t length)
  49 {
  50         while (length--) {
  51                 __asm__ __volatile__(
  52                         ".byte 0xf2, 0xf, 0x38, 0xf0, 0xf1"
  53                         :"=S"(crc)
  54                         :"0"(crc), "c"(*data)
  55                 );
  56                 data++;
  57         }
  58
  59         return crc;
  60 }
  61
  62 static u32 __pure crc32c_intel_le_hw(u32 crc, unsigned char const *p, size_t len)
  63 {
  64         unsigned int iquotient = len / SCALE_F;
  65         unsigned int iremainder = len % SCALE_F;
  66         unsigned long *ptmp = (unsigned long *)p;
  67
  68         while (iquotient--) {
  69                 __asm__ __volatile__(
  70                         ".byte 0xf2, " REX_PRE "0xf, 0x38, 0xf1, 0xf1;"
  71                         :"=S"(crc)
  72                         :"0"(crc), "c"(*ptmp)
  73                 );
  74                 ptmp++;
  75         }
  76
  77         if (iremainder)
  78                 crc = crc32c_intel_le_hw_byte(crc, (unsigned char *)ptmp,
  79                                  iremainder);
  80
  81         return crc;
  82 }
  83
  84 /*
  85  * Setting the seed allows arbitrary accumulators and flexible XOR policy
  86  * If your algorithm starts with ~0, then XOR with ~0 before you set
  87  * the seed.
  88  */
  89 static int crc32c_intel_setkey(struct crypto_shash *hash, const u8 *key,
  90                         unsigned int keylen)
  91 {
  92         u32 *mctx = crypto_shash_ctx(hash);
  93
  94         if (keylen != sizeof(u32)) {
  95                 crypto_shash_set_flags(hash, CRYPTO_TFM_RES_BAD_KEY_LEN);
  96                 return -EINVAL;
  97         }
  98         *mctx = le32_to_cpup((__le32 *)key);
  99         return 0;
 100 }
 101
 102 static int crc32c_intel_init(struct shash_desc *desc)
 103 {
 104         u32 *mctx = crypto_shash_ctx(desc->tfm);
 105         u32 *crcp = shash_desc_ctx(desc);
 106
 107         *crcp = *mctx;
 108
 109         return 0;
 110 }
 111
 112 static int crc32c_intel_update(struct shash_desc *desc, const u8 *data,
 113                                unsigned int len)
 114 {
 115         u32 *crcp = shash_desc_ctx(desc);
 116
 117         *crcp = crc32c_intel_le_hw(*crcp, data, len);
 118         return 0;
 119 }
 120
 121 static int __crc32c_intel_finup(u32 *crcp, const u8 *data, unsigned int len,
 122                                 u8 *out)
 123 {
 124         *(__le32 *)out = ~cpu_to_le32(crc32c_intel_le_hw(*crcp, data, len));
 125         return 0;
 126 }
 127
 128 static int crc32c_intel_finup(struct shash_desc *desc, const u8 *data,
 129                               unsigned int len, u8 *out)
 130 {
 131         return __crc32c_intel_finup(shash_desc_ctx(desc), data, len, out);
 132 }
 133
 134 static int crc32c_intel_final(struct shash_desc *desc, u8 *out)
 135 {
 136         u32 *crcp = shash_desc_ctx(desc);
 137
 138         *(__le32 *)out = ~cpu_to_le32p(crcp);
 139         return 0;
 140 }
 141
 142 static int crc32c_intel_digest(struct shash_desc *desc, const u8 *data,
 143                                unsigned int len, u8 *out)
 144 {
 145         return __crc32c_intel_finup(crypto_shash_ctx(desc->tfm), data, len,
 146                                     out);
 147 }
 148
 149 static int crc32c_intel_cra_init(struct crypto_tfm *tfm)
 150 {
 151         u32 *key = crypto_tfm_ctx(tfm);
 152
 153         *key = ~0;
 154
 155         return 0;
 156 }
 157
 158 #ifdef CONFIG_X86_64
 159 static int crc32c_pcl_intel_update(struct shash_desc *desc, const u8 *data,
 160                                unsigned int len)
 161 {
 162         u32 *crcp = shash_desc_ctx(desc);
 163
 164         /*
 165          * use faster PCL version if datasize is large enough to
 166          * overcome kernel fpu state save/restore overhead
 167          */
 168         if (len >= CRC32C_PCL_BREAKEVEN && crypto_simd_usable()) {
 169                 kernel_fpu_begin();
 170                 *crcp = crc_pcl(data, len, *crcp);
 171                 kernel_fpu_end();
 172         } else
 173                 *crcp = crc32c_intel_le_hw(*crcp, data, len);
 174         return 0;
 175 }
 176
 177 static int __crc32c_pcl_intel_finup(u32 *crcp, const u8 *data, unsigned int len,
 178                                 u8 *out)
 179 {
 180         if (len >= CRC32C_PCL_BREAKEVEN && crypto_simd_usable()) {
 181                 kernel_fpu_begin();
 182                 *(__le32 *)out = ~cpu_to_le32(crc_pcl(data, len, *crcp));
 183                 kernel_fpu_end();
 184         } else
 185                 *(__le32 *)out =
 186                         ~cpu_to_le32(crc32c_intel_le_hw(*crcp, data, len));
 187         return 0;
 188 }
 189
 190 static int crc32c_pcl_intel_finup(struct shash_desc *desc, const u8 *data,
 191                               unsigned int len, u8 *out)
 192 {
 193         return __crc32c_pcl_intel_finup(shash_desc_ctx(desc), data, len, out);
 194 }
 195
 196 static int crc32c_pcl_intel_digest(struct shash_desc *desc, const u8 *data,
 197                                unsigned int len, u8 *out)
 198 {
 199         return __crc32c_pcl_intel_finup(crypto_shash_ctx(desc->tfm), data, len,
 200                                     out);
 201 }
 202 #endif /* CONFIG_X86_64 */
 203
 204 static struct shash_alg alg = {
 205         .setkey                 =       crc32c_intel_setkey,
 206         .init                   =       crc32c_intel_init,
 207         .update                 =       crc32c_intel_update,
 208         .final                  =       crc32c_intel_final,
 209         .finup                  =       crc32c_intel_finup,
 210         .digest                 =       crc32c_intel_digest,
 211         .descsize               =       sizeof(u32),
 212         .digestsize             =       CHKSUM_DIGEST_SIZE,
 213         .base                   =       {
 214                 .cra_name               =       "crc32c",
 215                 .cra_driver_name        =       "crc32c-intel",
 216                 .cra_priority           =       200,
 217                 .cra_flags              =       CRYPTO_ALG_OPTIONAL_KEY,
 218                 .cra_blocksize          =       CHKSUM_BLOCK_SIZE,
 219                 .cra_ctxsize            =       sizeof(u32),
 220                 .cra_module             =       THIS_MODULE,
 221                 .cra_init               =       crc32c_intel_cra_init,
 222         }
 223 };
 224
 225 static const struct x86_cpu_id crc32c_cpu_id[] = {
 226         X86_FEATURE_MATCH(X86_FEATURE_XMM4_2),
 227         {}
 228 };
 229 MODULE_DEVICE_TABLE(x86cpu, crc32c_cpu_id);
 230
 231 static int __init crc32c_intel_mod_init(void)
 232 {
 233         if (!x86_match_cpu(crc32c_cpu_id))
 234                 return -ENODEV;
 235 #ifdef CONFIG_X86_64
 236         if (boot_cpu_has(X86_FEATURE_PCLMULQDQ)) {
 237                 alg.update = crc32c_pcl_intel_update;
 238                 alg.finup = crc32c_pcl_intel_finup;
 239                 alg.digest = crc32c_pcl_intel_digest;
 240         }
 241 #endif
 242         return crypto_register_shash(&alg);
 243 }
 244
 245 static void __exit crc32c_intel_mod_fini(void)
 246 {
 247         crypto_unregister_shash(&alg);
 248 }
 249
 250 module_init(crc32c_intel_mod_init);
 251 module_exit(crc32c_intel_mod_fini);
 252
 253 MODULE_AUTHOR("Austin Zhang <austin.zhang@intel.com>, Kent Liu <kent.liu@intel.com>");
 254 MODULE_DESCRIPTION("CRC32c (Castagnoli) optimization using Intel Hardware.");
 255 MODULE_LICENSE("GPL");
 256
 257 MODULE_ALIAS_CRYPTO("crc32c");
 258 MODULE_ALIAS_CRYPTO("crc32c-intel");