crypto/ppccap.c

   1 /*
   2  * Copyright 2009-2021 The OpenSSL Project Authors. All Rights Reserved.
   3  *
   4  * Licensed under the Apache License 2.0 (the "License").  You may not use
   5  * this file except in compliance with the License.  You can obtain a copy
   6  * in the file LICENSE in the source distribution or at
   7  * https://www.openssl.org/source/license.html
   8  */
   9
  10 #include <stdio.h>
  11 #include <stdlib.h>
  12 #include <string.h>
  13 #include <setjmp.h>
  14 #include <signal.h>
  15 #include <unistd.h>
  16 #if defined(__linux) || defined(_AIX)
  17 # include <sys/utsname.h>
  18 #endif
  19 #if defined(_AIX53)     /* defined even on post-5.3 */
  20 # include <sys/systemcfg.h>
  21 # if !defined(__power_set)
  22 #  define __power_set(a) (_system_configuration.implementation & (a))
  23 # endif
  24 #endif
  25 #if defined(__APPLE__) && defined(__MACH__)
  26 # include <sys/types.h>
  27 # include <sys/sysctl.h>
  28 #endif
  29 #include <openssl/crypto.h>
  30 #include <openssl/bn.h>
  31 #include <internal/cryptlib.h>
  32 #include <crypto/chacha.h>
  33 #include "bn/bn_local.h"
  34
  35 #include "ppc_arch.h"
  36
  37 unsigned int OPENSSL_ppccap_P = 0;
  38
  39 static sigset_t all_masked;
  40
  41
  42 #ifdef OPENSSL_BN_ASM_MONT
  43 int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
  44                 const BN_ULONG *np, const BN_ULONG *n0, int num)
  45 {
  46     int bn_mul_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
  47                         const BN_ULONG *np, const BN_ULONG *n0, int num);
  48     int bn_mul4x_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
  49                           const BN_ULONG *np, const BN_ULONG *n0, int num);
  50
  51     if (num < 4)
  52         return 0;
  53
  54     if ((num & 3) == 0)
  55         return bn_mul4x_mont_int(rp, ap, bp, np, n0, num);
  56
  57     /*
  58      * There used to be [optional] call to bn_mul_mont_fpu64 here,
  59      * but above subroutine is faster on contemporary processors.
  60      * Formulation means that there might be old processors where
  61      * FPU code path would be faster, POWER6 perhaps, but there was
  62      * no opportunity to figure it out...
  63      */
  64
  65     return bn_mul_mont_int(rp, ap, bp, np, n0, num);
  66 }
  67 #endif
  68 void sha256_block_p8(void *ctx, const void *inp, size_t len);
  69 void sha256_block_ppc(void *ctx, const void *inp, size_t len);
  70 void sha256_block_data_order(void *ctx, const void *inp, size_t len);
  71 void sha256_block_data_order(void *ctx, const void *inp, size_t len)
  72 {
  73     OPENSSL_ppccap_P & PPC_CRYPTO207 ? sha256_block_p8(ctx, inp, len) :
  74         sha256_block_ppc(ctx, inp, len);
  75 }
  76
  77 void sha512_block_p8(void *ctx, const void *inp, size_t len);
  78 void sha512_block_ppc(void *ctx, const void *inp, size_t len);
  79 void sha512_block_data_order(void *ctx, const void *inp, size_t len);
  80 void sha512_block_data_order(void *ctx, const void *inp, size_t len)
  81 {
  82     OPENSSL_ppccap_P & PPC_CRYPTO207 ? sha512_block_p8(ctx, inp, len) :
  83         sha512_block_ppc(ctx, inp, len);
  84 }
  85
  86 #ifndef FIPS_MODULE
  87 # ifndef OPENSSL_NO_CHACHA
  88 void ChaCha20_ctr32_int(unsigned char *out, const unsigned char *inp,
  89                         size_t len, const unsigned int key[8],
  90                         const unsigned int counter[4]);
  91 void ChaCha20_ctr32_vmx(unsigned char *out, const unsigned char *inp,
  92                         size_t len, const unsigned int key[8],
  93                         const unsigned int counter[4]);
  94 void ChaCha20_ctr32_vsx(unsigned char *out, const unsigned char *inp,
  95                         size_t len, const unsigned int key[8],
  96                         const unsigned int counter[4]);
  97 void ChaCha20_ctr32(unsigned char *out, const unsigned char *inp,
  98                     size_t len, const unsigned int key[8],
  99                     const unsigned int counter[4])
 100 {
 101     OPENSSL_ppccap_P & PPC_CRYPTO207
 102         ? ChaCha20_ctr32_vsx(out, inp, len, key, counter)
 103         : OPENSSL_ppccap_P & PPC_ALTIVEC
 104             ? ChaCha20_ctr32_vmx(out, inp, len, key, counter)
 105             : ChaCha20_ctr32_int(out, inp, len, key, counter);
 106 }
 107 # endif
 108
 109 # ifndef OPENSSL_NO_POLY1305
 110 void poly1305_init_int(void *ctx, const unsigned char key[16]);
 111 void poly1305_blocks(void *ctx, const unsigned char *inp, size_t len,
 112                          unsigned int padbit);
 113 void poly1305_emit(void *ctx, unsigned char mac[16],
 114                        const unsigned int nonce[4]);
 115 void poly1305_init_fpu(void *ctx, const unsigned char key[16]);
 116 void poly1305_blocks_fpu(void *ctx, const unsigned char *inp, size_t len,
 117                          unsigned int padbit);
 118 void poly1305_emit_fpu(void *ctx, unsigned char mac[16],
 119                        const unsigned int nonce[4]);
 120 void poly1305_init_vsx(void *ctx, const unsigned char key[16]);
 121 void poly1305_blocks_vsx(void *ctx, const unsigned char *inp, size_t len,
 122                          unsigned int padbit);
 123 void poly1305_emit_vsx(void *ctx, unsigned char mac[16],
 124                        const unsigned int nonce[4]);
 125 int poly1305_init(void *ctx, const unsigned char key[16], void *func[2]);
 126 int poly1305_init(void *ctx, const unsigned char key[16], void *func[2])
 127 {
 128     if (OPENSSL_ppccap_P & PPC_CRYPTO207) {
 129         poly1305_init_int(ctx, key);
 130         func[0] = (void*)(uintptr_t)poly1305_blocks_vsx;
 131         func[1] = (void*)(uintptr_t)poly1305_emit;
 132     } else if (sizeof(size_t) == 4 && (OPENSSL_ppccap_P & PPC_FPU)) {
 133         poly1305_init_fpu(ctx, key);
 134         func[0] = (void*)(uintptr_t)poly1305_blocks_fpu;
 135         func[1] = (void*)(uintptr_t)poly1305_emit_fpu;
 136     } else {
 137         poly1305_init_int(ctx, key);
 138         func[0] = (void*)(uintptr_t)poly1305_blocks;
 139         func[1] = (void*)(uintptr_t)poly1305_emit;
 140     }
 141     return 1;
 142 }
 143 # endif
 144 #endif /* FIPS_MODULE */
 145
 146 #ifdef ECP_NISTZ256_ASM
 147 void ecp_nistz256_mul_mont(unsigned long res[4], const unsigned long a[4],
 148                            const unsigned long b[4]);
 149
 150 void ecp_nistz256_to_mont(unsigned long res[4], const unsigned long in[4]);
 151 void ecp_nistz256_to_mont(unsigned long res[4], const unsigned long in[4])
 152 {
 153     static const unsigned long RR[] = { 0x0000000000000003U,
 154                                         0xfffffffbffffffffU,
 155                                         0xfffffffffffffffeU,
 156                                         0x00000004fffffffdU };
 157
 158     ecp_nistz256_mul_mont(res, in, RR);
 159 }
 160
 161 void ecp_nistz256_from_mont(unsigned long res[4], const unsigned long in[4]);
 162 void ecp_nistz256_from_mont(unsigned long res[4], const unsigned long in[4])
 163 {
 164     static const unsigned long one[] = { 1, 0, 0, 0 };
 165
 166     ecp_nistz256_mul_mont(res, in, one);
 167 }
 168 #endif
 169
 170 static sigjmp_buf ill_jmp;
 171 static void ill_handler(int sig)
 172 {
 173     siglongjmp(ill_jmp, sig);
 174 }
 175
 176 void OPENSSL_fpu_probe(void);
 177 void OPENSSL_ppc64_probe(void);
 178 void OPENSSL_altivec_probe(void);
 179 void OPENSSL_crypto207_probe(void);
 180 void OPENSSL_madd300_probe(void);
 181
 182 long OPENSSL_rdtsc_mftb(void);
 183 long OPENSSL_rdtsc_mfspr268(void);
 184
 185 uint32_t OPENSSL_rdtsc(void)
 186 {
 187     if (OPENSSL_ppccap_P & PPC_MFTB)
 188         return OPENSSL_rdtsc_mftb();
 189     else if (OPENSSL_ppccap_P & PPC_MFSPR268)
 190         return OPENSSL_rdtsc_mfspr268();
 191     else
 192         return 0;
 193 }
 194
 195 size_t OPENSSL_instrument_bus_mftb(unsigned int *, size_t);
 196 size_t OPENSSL_instrument_bus_mfspr268(unsigned int *, size_t);
 197
 198 size_t OPENSSL_instrument_bus(unsigned int *out, size_t cnt)
 199 {
 200     if (OPENSSL_ppccap_P & PPC_MFTB)
 201         return OPENSSL_instrument_bus_mftb(out, cnt);
 202     else if (OPENSSL_ppccap_P & PPC_MFSPR268)
 203         return OPENSSL_instrument_bus_mfspr268(out, cnt);
 204     else
 205         return 0;
 206 }
 207
 208 size_t OPENSSL_instrument_bus2_mftb(unsigned int *, size_t, size_t);
 209 size_t OPENSSL_instrument_bus2_mfspr268(unsigned int *, size_t, size_t);
 210
 211 size_t OPENSSL_instrument_bus2(unsigned int *out, size_t cnt, size_t max)
 212 {
 213     if (OPENSSL_ppccap_P & PPC_MFTB)
 214         return OPENSSL_instrument_bus2_mftb(out, cnt, max);
 215     else if (OPENSSL_ppccap_P & PPC_MFSPR268)
 216         return OPENSSL_instrument_bus2_mfspr268(out, cnt, max);
 217     else
 218         return 0;
 219 }
 220
 221 #if defined(__GLIBC__) && defined(__GLIBC_PREREQ)
 222 # if __GLIBC_PREREQ(2, 16)
 223 #  include <sys/auxv.h>
 224 #  define OSSL_IMPLEMENT_GETAUXVAL
 225 # endif
 226 #endif
 227
 228 #if defined(__FreeBSD__)
 229 # include <sys/param.h>
 230 # if __FreeBSD_version >= 1200000
 231 #  include <sys/auxv.h>
 232 #  define OSSL_IMPLEMENT_GETAUXVAL
 233
 234 static unsigned long getauxval(unsigned long key)
 235 {
 236   unsigned long val = 0ul;
 237
 238   if (elf_aux_info((int)key, &val, sizeof(val)) != 0)
 239     return 0ul;
 240
 241   return val;
 242 }
 243 # endif
 244 #endif
 245
 246 /* I wish <sys/auxv.h> was universally available */
 247 #define HWCAP                   16      /* AT_HWCAP */
 248 #define HWCAP_PPC64             (1U << 30)
 249 #define HWCAP_ALTIVEC           (1U << 28)
 250 #define HWCAP_FPU               (1U << 27)
 251 #define HWCAP_POWER6_EXT        (1U << 9)
 252 #define HWCAP_VSX               (1U << 7)
 253
 254 #define HWCAP2                  26      /* AT_HWCAP2 */
 255 #define HWCAP_VEC_CRYPTO        (1U << 25)
 256 #define HWCAP_ARCH_3_00         (1U << 23)
 257
 258 # if defined(__GNUC__) && __GNUC__>=2
 259 __attribute__ ((constructor))
 260 # endif
 261 void OPENSSL_cpuid_setup(void)
 262 {
 263     char *e;
 264     struct sigaction ill_oact, ill_act;
 265     sigset_t oset;
 266     static int trigger = 0;
 267
 268     if (trigger)
 269         return;
 270     trigger = 1;
 271
 272     if ((e = getenv("OPENSSL_ppccap"))) {
 273         OPENSSL_ppccap_P = strtoul(e, NULL, 0);
 274         return;
 275     }
 276
 277     OPENSSL_ppccap_P = 0;
 278
 279 #if defined(_AIX)
 280     OPENSSL_ppccap_P |= PPC_FPU;
 281
 282     if (sizeof(size_t) == 4) {
 283         struct utsname uts;
 284 # if defined(_SC_AIX_KERNEL_BITMODE)
 285         if (sysconf(_SC_AIX_KERNEL_BITMODE) != 64)
 286             return;
 287 # endif
 288         if (uname(&uts) != 0 || atoi(uts.version) < 6)
 289             return;
 290     }
 291
 292 # if defined(__power_set)
 293     /*
 294      * Value used in __power_set is a single-bit 1<<n one denoting
 295      * specific processor class. Incidentally 0xffffffff<<n can be
 296      * used to denote specific processor and its successors.
 297      */
 298     if (sizeof(size_t) == 4) {
 299         /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 300         if (__power_set(0xffffffffU<<13))       /* POWER5 and later */
 301             OPENSSL_ppccap_P |= PPC_FPU64;
 302     } else {
 303         /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 304         if (__power_set(0x1U<<14))              /* POWER6 */
 305             OPENSSL_ppccap_P |= PPC_FPU64;
 306     }
 307
 308     if (__power_set(0xffffffffU<<14))           /* POWER6 and later */
 309         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 310
 311     if (__power_set(0xffffffffU<<16))           /* POWER8 and later */
 312         OPENSSL_ppccap_P |= PPC_CRYPTO207;
 313
 314     if (__power_set(0xffffffffU<<17))           /* POWER9 and later */
 315         OPENSSL_ppccap_P |= PPC_MADD300;
 316
 317     return;
 318 # endif
 319 #endif
 320
 321 #if defined(__APPLE__) && defined(__MACH__)
 322     OPENSSL_ppccap_P |= PPC_FPU;
 323
 324     {
 325         int val;
 326         size_t len = sizeof(val);
 327
 328         if (sysctlbyname("hw.optional.64bitops", &val, &len, NULL, 0) == 0) {
 329             if (val)
 330                 OPENSSL_ppccap_P |= PPC_FPU64;
 331         }
 332
 333         len = sizeof(val);
 334         if (sysctlbyname("hw.optional.altivec", &val, &len, NULL, 0) == 0) {
 335             if (val)
 336                 OPENSSL_ppccap_P |= PPC_ALTIVEC;
 337         }
 338
 339         return;
 340     }
 341 #endif
 342
 343 #ifdef OSSL_IMPLEMENT_GETAUXVAL
 344     {
 345         unsigned long hwcap = getauxval(HWCAP);
 346         unsigned long hwcap2 = getauxval(HWCAP2);
 347
 348         if (hwcap & HWCAP_FPU) {
 349             OPENSSL_ppccap_P |= PPC_FPU;
 350
 351             if (sizeof(size_t) == 4) {
 352                 /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 353                 if (hwcap & HWCAP_PPC64)
 354                     OPENSSL_ppccap_P |= PPC_FPU64;
 355             } else {
 356                 /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 357                 if (hwcap & HWCAP_POWER6_EXT)
 358                     OPENSSL_ppccap_P |= PPC_FPU64;
 359             }
 360         }
 361
 362         if (hwcap & HWCAP_ALTIVEC) {
 363             OPENSSL_ppccap_P |= PPC_ALTIVEC;
 364
 365             if ((hwcap & HWCAP_VSX) && (hwcap2 & HWCAP_VEC_CRYPTO))
 366                 OPENSSL_ppccap_P |= PPC_CRYPTO207;
 367         }
 368
 369         if (hwcap2 & HWCAP_ARCH_3_00) {
 370             OPENSSL_ppccap_P |= PPC_MADD300;
 371         }
 372     }
 373 #endif
 374
 375     sigfillset(&all_masked);
 376     sigdelset(&all_masked, SIGILL);
 377     sigdelset(&all_masked, SIGTRAP);
 378 #ifdef SIGEMT
 379     sigdelset(&all_masked, SIGEMT);
 380 #endif
 381     sigdelset(&all_masked, SIGFPE);
 382     sigdelset(&all_masked, SIGBUS);
 383     sigdelset(&all_masked, SIGSEGV);
 384
 385     memset(&ill_act, 0, sizeof(ill_act));
 386     ill_act.sa_handler = ill_handler;
 387     ill_act.sa_mask = all_masked;
 388
 389     sigprocmask(SIG_SETMASK, &ill_act.sa_mask, &oset);
 390     sigaction(SIGILL, &ill_act, &ill_oact);
 391
 392 #ifndef OSSL_IMPLEMENT_GETAUXVAL
 393     if (sigsetjmp(ill_jmp,1) == 0) {
 394         OPENSSL_fpu_probe();
 395         OPENSSL_ppccap_P |= PPC_FPU;
 396
 397         if (sizeof(size_t) == 4) {
 398 # ifdef __linux
 399             struct utsname uts;
 400             if (uname(&uts) == 0 && strcmp(uts.machine, "ppc64") == 0)
 401 # endif
 402                 if (sigsetjmp(ill_jmp, 1) == 0) {
 403                     OPENSSL_ppc64_probe();
 404                     OPENSSL_ppccap_P |= PPC_FPU64;
 405                 }
 406         } else {
 407             /*
 408              * Wanted code detecting POWER6 CPU and setting PPC_FPU64
 409              */
 410         }
 411     }
 412
 413     if (sigsetjmp(ill_jmp, 1) == 0) {
 414         OPENSSL_altivec_probe();
 415         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 416         if (sigsetjmp(ill_jmp, 1) == 0) {
 417             OPENSSL_crypto207_probe();
 418             OPENSSL_ppccap_P |= PPC_CRYPTO207;
 419         }
 420     }
 421
 422     if (sigsetjmp(ill_jmp, 1) == 0) {
 423         OPENSSL_madd300_probe();
 424         OPENSSL_ppccap_P |= PPC_MADD300;
 425     }
 426 #endif
 427
 428     if (sigsetjmp(ill_jmp, 1) == 0) {
 429         OPENSSL_rdtsc_mftb();
 430         OPENSSL_ppccap_P |= PPC_MFTB;
 431     } else if (sigsetjmp(ill_jmp, 1) == 0) {
 432         OPENSSL_rdtsc_mfspr268();
 433         OPENSSL_ppccap_P |= PPC_MFSPR268;
 434     }
 435
 436     sigaction(SIGILL, &ill_oact, NULL);
 437     sigprocmask(SIG_SETMASK, &oset, NULL);
 438 }