crypto/ppccap.c

   1 /*
   2  * Copyright 2009-2018 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 /*
  87  * TODO(3.0): Temporarily disabled some assembler that hasn't been brought into
  88  * the FIPS module yet.
  89  */
  90 #ifndef FIPS_MODE
  91 # ifndef OPENSSL_NO_CHACHA
  92 void ChaCha20_ctr32_int(unsigned char *out, const unsigned char *inp,
  93                         size_t len, const unsigned int key[8],
  94                         const unsigned int counter[4]);
  95 void ChaCha20_ctr32_vmx(unsigned char *out, const unsigned char *inp,
  96                         size_t len, const unsigned int key[8],
  97                         const unsigned int counter[4]);
  98 void ChaCha20_ctr32_vsx(unsigned char *out, const unsigned char *inp,
  99                         size_t len, const unsigned int key[8],
 100                         const unsigned int counter[4]);
 101 void ChaCha20_ctr32(unsigned char *out, const unsigned char *inp,
 102                     size_t len, const unsigned int key[8],
 103                     const unsigned int counter[4])
 104 {
 105     OPENSSL_ppccap_P & PPC_CRYPTO207
 106         ? ChaCha20_ctr32_vsx(out, inp, len, key, counter)
 107         : OPENSSL_ppccap_P & PPC_ALTIVEC
 108             ? ChaCha20_ctr32_vmx(out, inp, len, key, counter)
 109             : ChaCha20_ctr32_int(out, inp, len, key, counter);
 110 }
 111 # endif
 112
 113 # ifndef OPENSSL_NO_POLY1305
 114 void poly1305_init_int(void *ctx, const unsigned char key[16]);
 115 void poly1305_blocks(void *ctx, const unsigned char *inp, size_t len,
 116                          unsigned int padbit);
 117 void poly1305_emit(void *ctx, unsigned char mac[16],
 118                        const unsigned int nonce[4]);
 119 void poly1305_init_fpu(void *ctx, const unsigned char key[16]);
 120 void poly1305_blocks_fpu(void *ctx, const unsigned char *inp, size_t len,
 121                          unsigned int padbit);
 122 void poly1305_emit_fpu(void *ctx, unsigned char mac[16],
 123                        const unsigned int nonce[4]);
 124 void poly1305_init_vsx(void *ctx, const unsigned char key[16]);
 125 void poly1305_blocks_vsx(void *ctx, const unsigned char *inp, size_t len,
 126                          unsigned int padbit);
 127 void poly1305_emit_vsx(void *ctx, unsigned char mac[16],
 128                        const unsigned int nonce[4]);
 129 int poly1305_init(void *ctx, const unsigned char key[16], void *func[2]);
 130 int poly1305_init(void *ctx, const unsigned char key[16], void *func[2])
 131 {
 132     if (OPENSSL_ppccap_P & PPC_CRYPTO207) {
 133         poly1305_init_int(ctx, key);
 134         func[0] = (void*)(uintptr_t)poly1305_blocks_vsx;
 135         func[1] = (void*)(uintptr_t)poly1305_emit;
 136     } else if (sizeof(size_t) == 4 && (OPENSSL_ppccap_P & PPC_FPU)) {
 137         poly1305_init_fpu(ctx, key);
 138         func[0] = (void*)(uintptr_t)poly1305_blocks_fpu;
 139         func[1] = (void*)(uintptr_t)poly1305_emit_fpu;
 140     } else {
 141         poly1305_init_int(ctx, key);
 142         func[0] = (void*)(uintptr_t)poly1305_blocks;
 143         func[1] = (void*)(uintptr_t)poly1305_emit;
 144     }
 145     return 1;
 146 }
 147 # endif
 148 #endif /* FIPS_MODE */
 149
 150 #ifdef ECP_NISTZ256_ASM
 151 void ecp_nistz256_mul_mont(unsigned long res[4], const unsigned long a[4],
 152                            const unsigned long b[4]);
 153
 154 void ecp_nistz256_to_mont(unsigned long res[4], const unsigned long in[4]);
 155 void ecp_nistz256_to_mont(unsigned long res[4], const unsigned long in[4])
 156 {
 157     static const unsigned long RR[] = { 0x0000000000000003U,
 158                                         0xfffffffbffffffffU,
 159                                         0xfffffffffffffffeU,
 160                                         0x00000004fffffffdU };
 161
 162     ecp_nistz256_mul_mont(res, in, RR);
 163 }
 164
 165 void ecp_nistz256_from_mont(unsigned long res[4], const unsigned long in[4]);
 166 void ecp_nistz256_from_mont(unsigned long res[4], const unsigned long in[4])
 167 {
 168     static const unsigned long one[] = { 1, 0, 0, 0 };
 169
 170     ecp_nistz256_mul_mont(res, in, one);
 171 }
 172 #endif
 173
 174 static sigjmp_buf ill_jmp;
 175 static void ill_handler(int sig)
 176 {
 177     siglongjmp(ill_jmp, sig);
 178 }
 179
 180 void OPENSSL_fpu_probe(void);
 181 void OPENSSL_ppc64_probe(void);
 182 void OPENSSL_altivec_probe(void);
 183 void OPENSSL_crypto207_probe(void);
 184 void OPENSSL_madd300_probe(void);
 185
 186 long OPENSSL_rdtsc_mftb(void);
 187 long OPENSSL_rdtsc_mfspr268(void);
 188
 189 uint32_t OPENSSL_rdtsc(void)
 190 {
 191     if (OPENSSL_ppccap_P & PPC_MFTB)
 192         return OPENSSL_rdtsc_mftb();
 193     else if (OPENSSL_ppccap_P & PPC_MFSPR268)
 194         return OPENSSL_rdtsc_mfspr268();
 195     else
 196         return 0;
 197 }
 198
 199 size_t OPENSSL_instrument_bus_mftb(unsigned int *, size_t);
 200 size_t OPENSSL_instrument_bus_mfspr268(unsigned int *, size_t);
 201
 202 size_t OPENSSL_instrument_bus(unsigned int *out, size_t cnt)
 203 {
 204     if (OPENSSL_ppccap_P & PPC_MFTB)
 205         return OPENSSL_instrument_bus_mftb(out, cnt);
 206     else if (OPENSSL_ppccap_P & PPC_MFSPR268)
 207         return OPENSSL_instrument_bus_mfspr268(out, cnt);
 208     else
 209         return 0;
 210 }
 211
 212 size_t OPENSSL_instrument_bus2_mftb(unsigned int *, size_t, size_t);
 213 size_t OPENSSL_instrument_bus2_mfspr268(unsigned int *, size_t, size_t);
 214
 215 size_t OPENSSL_instrument_bus2(unsigned int *out, size_t cnt, size_t max)
 216 {
 217     if (OPENSSL_ppccap_P & PPC_MFTB)
 218         return OPENSSL_instrument_bus2_mftb(out, cnt, max);
 219     else if (OPENSSL_ppccap_P & PPC_MFSPR268)
 220         return OPENSSL_instrument_bus2_mfspr268(out, cnt, max);
 221     else
 222         return 0;
 223 }
 224
 225 #if defined(__GLIBC__) && defined(__GLIBC_PREREQ)
 226 # if __GLIBC_PREREQ(2, 16)
 227 #  include <sys/auxv.h>
 228 #  define OSSL_IMPLEMENT_GETAUXVAL
 229 # endif
 230 #endif
 231
 232 /* I wish <sys/auxv.h> was universally available */
 233 #define HWCAP                   16      /* AT_HWCAP */
 234 #define HWCAP_PPC64             (1U << 30)
 235 #define HWCAP_ALTIVEC           (1U << 28)
 236 #define HWCAP_FPU               (1U << 27)
 237 #define HWCAP_POWER6_EXT        (1U << 9)
 238 #define HWCAP_VSX               (1U << 7)
 239
 240 #define HWCAP2                  26      /* AT_HWCAP2 */
 241 #define HWCAP_VEC_CRYPTO        (1U << 25)
 242 #define HWCAP_ARCH_3_00         (1U << 23)
 243
 244 # if defined(__GNUC__) && __GNUC__>=2
 245 __attribute__ ((constructor))
 246 # endif
 247 void OPENSSL_cpuid_setup(void)
 248 {
 249     char *e;
 250     struct sigaction ill_oact, ill_act;
 251     sigset_t oset;
 252     static int trigger = 0;
 253
 254     if (trigger)
 255         return;
 256     trigger = 1;
 257
 258     if ((e = getenv("OPENSSL_ppccap"))) {
 259         OPENSSL_ppccap_P = strtoul(e, NULL, 0);
 260         return;
 261     }
 262
 263     OPENSSL_ppccap_P = 0;
 264
 265 #if defined(_AIX)
 266     OPENSSL_ppccap_P |= PPC_FPU;
 267
 268     if (sizeof(size_t) == 4) {
 269         struct utsname uts;
 270 # if defined(_SC_AIX_KERNEL_BITMODE)
 271         if (sysconf(_SC_AIX_KERNEL_BITMODE) != 64)
 272             return;
 273 # endif
 274         if (uname(&uts) != 0 || atoi(uts.version) < 6)
 275             return;
 276     }
 277
 278 # if defined(__power_set)
 279     /*
 280      * Value used in __power_set is a single-bit 1<<n one denoting
 281      * specific processor class. Incidentally 0xffffffff<<n can be
 282      * used to denote specific processor and its successors.
 283      */
 284     if (sizeof(size_t) == 4) {
 285         /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 286         if (__power_set(0xffffffffU<<13))       /* POWER5 and later */
 287             OPENSSL_ppccap_P |= PPC_FPU64;
 288     } else {
 289         /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 290         if (__power_set(0x1U<<14))              /* POWER6 */
 291             OPENSSL_ppccap_P |= PPC_FPU64;
 292     }
 293
 294     if (__power_set(0xffffffffU<<14))           /* POWER6 and later */
 295         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 296
 297     if (__power_set(0xffffffffU<<16))           /* POWER8 and later */
 298         OPENSSL_ppccap_P |= PPC_CRYPTO207;
 299
 300     if (__power_set(0xffffffffU<<17))           /* POWER9 and later */
 301         OPENSSL_ppccap_P |= PPC_MADD300;
 302
 303     return;
 304 # endif
 305 #endif
 306
 307 #if defined(__APPLE__) && defined(__MACH__)
 308     OPENSSL_ppccap_P |= PPC_FPU;
 309
 310     {
 311         int val;
 312         size_t len = sizeof(val);
 313
 314         if (sysctlbyname("hw.optional.64bitops", &val, &len, NULL, 0) == 0) {
 315             if (val)
 316                 OPENSSL_ppccap_P |= PPC_FPU64;
 317         }
 318
 319         len = sizeof(val);
 320         if (sysctlbyname("hw.optional.altivec", &val, &len, NULL, 0) == 0) {
 321             if (val)
 322                 OPENSSL_ppccap_P |= PPC_ALTIVEC;
 323         }
 324
 325         return;
 326     }
 327 #endif
 328
 329 #ifdef OSSL_IMPLEMENT_GETAUXVAL
 330     {
 331         unsigned long hwcap = getauxval(HWCAP);
 332         unsigned long hwcap2 = getauxval(HWCAP2);
 333
 334         if (hwcap & HWCAP_FPU) {
 335             OPENSSL_ppccap_P |= PPC_FPU;
 336
 337             if (sizeof(size_t) == 4) {
 338                 /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 339                 if (hwcap & HWCAP_PPC64)
 340                     OPENSSL_ppccap_P |= PPC_FPU64;
 341             } else {
 342                 /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 343                 if (hwcap & HWCAP_POWER6_EXT)
 344                     OPENSSL_ppccap_P |= PPC_FPU64;
 345             }
 346         }
 347
 348         if (hwcap & HWCAP_ALTIVEC) {
 349             OPENSSL_ppccap_P |= PPC_ALTIVEC;
 350
 351             if ((hwcap & HWCAP_VSX) && (hwcap2 & HWCAP_VEC_CRYPTO))
 352                 OPENSSL_ppccap_P |= PPC_CRYPTO207;
 353         }
 354
 355         if (hwcap2 & HWCAP_ARCH_3_00) {
 356             OPENSSL_ppccap_P |= PPC_MADD300;
 357         }
 358     }
 359 #endif
 360
 361     sigfillset(&all_masked);
 362     sigdelset(&all_masked, SIGILL);
 363     sigdelset(&all_masked, SIGTRAP);
 364 #ifdef SIGEMT
 365     sigdelset(&all_masked, SIGEMT);
 366 #endif
 367     sigdelset(&all_masked, SIGFPE);
 368     sigdelset(&all_masked, SIGBUS);
 369     sigdelset(&all_masked, SIGSEGV);
 370
 371     memset(&ill_act, 0, sizeof(ill_act));
 372     ill_act.sa_handler = ill_handler;
 373     ill_act.sa_mask = all_masked;
 374
 375     sigprocmask(SIG_SETMASK, &ill_act.sa_mask, &oset);
 376     sigaction(SIGILL, &ill_act, &ill_oact);
 377
 378 #ifndef OSSL_IMPLEMENT_GETAUXVAL
 379     if (sigsetjmp(ill_jmp,1) == 0) {
 380         OPENSSL_fpu_probe();
 381         OPENSSL_ppccap_P |= PPC_FPU;
 382
 383         if (sizeof(size_t) == 4) {
 384 # ifdef __linux
 385             struct utsname uts;
 386             if (uname(&uts) == 0 && strcmp(uts.machine, "ppc64") == 0)
 387 # endif
 388                 if (sigsetjmp(ill_jmp, 1) == 0) {
 389                     OPENSSL_ppc64_probe();
 390                     OPENSSL_ppccap_P |= PPC_FPU64;
 391                 }
 392         } else {
 393             /*
 394              * Wanted code detecting POWER6 CPU and setting PPC_FPU64
 395              */
 396         }
 397     }
 398
 399     if (sigsetjmp(ill_jmp, 1) == 0) {
 400         OPENSSL_altivec_probe();
 401         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 402         if (sigsetjmp(ill_jmp, 1) == 0) {
 403             OPENSSL_crypto207_probe();
 404             OPENSSL_ppccap_P |= PPC_CRYPTO207;
 405         }
 406     }
 407
 408     if (sigsetjmp(ill_jmp, 1) == 0) {
 409         OPENSSL_madd300_probe();
 410         OPENSSL_ppccap_P |= PPC_MADD300;
 411     }
 412 #endif
 413
 414     if (sigsetjmp(ill_jmp, 1) == 0) {
 415         OPENSSL_rdtsc_mftb();
 416         OPENSSL_ppccap_P |= PPC_MFTB;
 417     } else if (sigsetjmp(ill_jmp, 1) == 0) {
 418         OPENSSL_rdtsc_mfspr268();
 419         OPENSSL_ppccap_P |= PPC_MFSPR268;
 420     }
 421
 422     sigaction(SIGILL, &ill_oact, NULL);
 423     sigprocmask(SIG_SETMASK, &oset, NULL);
 424 }