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