crypto/ppccap.c

   1 /*
   2  * Copyright 2009-2016 The OpenSSL Project Authors. All Rights Reserved.
   3  *
   4  * Licensed under the OpenSSL license (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 #include <openssl/crypto.h>
  26 #include <openssl/bn.h>
  27
  28 #include "ppc_arch.h"
  29
  30 unsigned int OPENSSL_ppccap_P = 0;
  31
  32 static sigset_t all_masked;
  33
  34 #ifdef OPENSSL_BN_ASM_MONT
  35 int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
  36                 const BN_ULONG *np, const BN_ULONG *n0, int num)
  37 {
  38     int bn_mul_mont_fpu64(BN_ULONG *rp, const BN_ULONG *ap,
  39                           const BN_ULONG *bp, const BN_ULONG *np,
  40                           const BN_ULONG *n0, int num);
  41     int bn_mul_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
  42                         const BN_ULONG *np, const BN_ULONG *n0, int num);
  43
  44     if (sizeof(size_t) == 4) {
  45 # if 1 || (defined(__APPLE__) && defined(__MACH__))
  46         if (num >= 8 && (num & 3) == 0 && (OPENSSL_ppccap_P & PPC_FPU64))
  47             return bn_mul_mont_fpu64(rp, ap, bp, np, n0, num);
  48 # else
  49         /*
  50          * boundary of 32 was experimentally determined on Linux 2.6.22,
  51          * might have to be adjusted on AIX...
  52          */
  53         if (num >= 32 && (num & 3) == 0 && (OPENSSL_ppccap_P & PPC_FPU64)) {
  54             sigset_t oset;
  55             int ret;
  56
  57             sigprocmask(SIG_SETMASK, &all_masked, &oset);
  58             ret = bn_mul_mont_fpu64(rp, ap, bp, np, n0, num);
  59             sigprocmask(SIG_SETMASK, &oset, NULL);
  60
  61             return ret;
  62         }
  63 # endif
  64     } else if ((OPENSSL_ppccap_P & PPC_FPU64))
  65         /*
  66          * this is a "must" on POWER6, but run-time detection is not
  67          * implemented yet...
  68          */
  69         return bn_mul_mont_fpu64(rp, ap, bp, np, n0, num);
  70
  71     return bn_mul_mont_int(rp, ap, bp, np, n0, num);
  72 }
  73 #endif
  74
  75 void sha256_block_p8(void *ctx, const void *inp, size_t len);
  76 void sha256_block_ppc(void *ctx, const void *inp, size_t len);
  77 void sha256_block_data_order(void *ctx, const void *inp, size_t len)
  78 {
  79     OPENSSL_ppccap_P & PPC_CRYPTO207 ? sha256_block_p8(ctx, inp, len) :
  80         sha256_block_ppc(ctx, inp, len);
  81 }
  82
  83 void sha512_block_p8(void *ctx, const void *inp, size_t len);
  84 void sha512_block_ppc(void *ctx, const void *inp, size_t len);
  85 void sha512_block_data_order(void *ctx, const void *inp, size_t len)
  86 {
  87     OPENSSL_ppccap_P & PPC_CRYPTO207 ? sha512_block_p8(ctx, inp, len) :
  88         sha512_block_ppc(ctx, inp, len);
  89 }
  90
  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(unsigned char *out, const unsigned char *inp,
  99                     size_t len, const unsigned int key[8],
 100                     const unsigned int counter[4])
 101 {
 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 int poly1305_init(void *ctx, const unsigned char key[16], void *func[2])
 120 {
 121     if (sizeof(size_t) == 4 && (OPENSSL_ppccap_P & PPC_FPU)) {
 122         poly1305_init_fpu(ctx, key);
 123         func[0] = poly1305_blocks_fpu;
 124         func[1] = poly1305_emit_fpu;
 125     } else {
 126         poly1305_init_int(ctx, key);
 127         func[0] = poly1305_blocks;
 128         func[1] = poly1305_emit;
 129     }
 130     return 1;
 131 }
 132 #endif
 133
 134 #ifdef ECP_NISTZ256_ASM
 135 void ecp_nistz256_mul_mont(unsigned long res[4], const unsigned long a[4],
 136                            const unsigned long b[4]);
 137
 138 void ecp_nistz256_to_mont(unsigned long res[4], const unsigned long in[4]);
 139 void ecp_nistz256_to_mont(unsigned long res[4], const unsigned long in[4])
 140 {
 141     static const unsigned long RR[] = { 0x0000000000000003U,
 142                                         0xfffffffbffffffffU,
 143                                         0xfffffffffffffffeU,
 144                                         0x00000004fffffffdU };
 145
 146     ecp_nistz256_mul_mont(res, in, RR);
 147 }
 148
 149 void ecp_nistz256_from_mont(unsigned long res[4], const unsigned long in[4]);
 150 void ecp_nistz256_from_mont(unsigned long res[4], const unsigned long in[4])
 151 {
 152     static const unsigned long one[] = { 1, 0, 0, 0 };
 153
 154     ecp_nistz256_mul_mont(res, in, one);
 155 }
 156 #endif
 157
 158 static sigjmp_buf ill_jmp;
 159 static void ill_handler(int sig)
 160 {
 161     siglongjmp(ill_jmp, sig);
 162 }
 163
 164 void OPENSSL_fpu_probe(void);
 165 void OPENSSL_ppc64_probe(void);
 166 void OPENSSL_altivec_probe(void);
 167 void OPENSSL_crypto207_probe(void);
 168 void OPENSSL_madd300_probe(void);
 169
 170 /*
 171  * Use a weak reference to getauxval() so we can use it if it is available
 172  * but don't break the build if it is not. Note that this is *link-time*
 173  * feature detection, not *run-time*. In other words if we link with
 174  * symbol present, it's expected to be present even at run-time.
 175  */
 176 #if defined(__GNUC__) && __GNUC__>=2 && defined(__ELF__)
 177 extern unsigned long getauxval(unsigned long type) __attribute__ ((weak));
 178 #else
 179 static unsigned long (*getauxval) (unsigned long) = NULL;
 180 #endif
 181
 182 /* I wish <sys/auxv.h> was universally available */
 183 #define HWCAP                   16      /* AT_HWCAP */
 184 #define HWCAP_PPC64             (1U << 30)
 185 #define HWCAP_ALTIVEC           (1U << 28)
 186 #define HWCAP_FPU               (1U << 27)
 187 #define HWCAP_POWER6_EXT        (1U << 9)
 188 #define HWCAP_VSX               (1U << 7)
 189
 190 #define HWCAP2                  26      /* AT_HWCAP2 */
 191 #define HWCAP_VEC_CRYPTO        (1U << 25)
 192 #define HWCAP_ARCH_3_00         (1U << 23)
 193
 194 # if defined(__GNUC__) && __GNUC__>=2
 195 __attribute__ ((constructor))
 196 # endif
 197 void OPENSSL_cpuid_setup(void)
 198 {
 199     char *e;
 200     struct sigaction ill_oact, ill_act;
 201     sigset_t oset;
 202     static int trigger = 0;
 203
 204     if (trigger)
 205         return;
 206     trigger = 1;
 207
 208     if ((e = getenv("OPENSSL_ppccap"))) {
 209         OPENSSL_ppccap_P = strtoul(e, NULL, 0);
 210         return;
 211     }
 212
 213     OPENSSL_ppccap_P = 0;
 214
 215 #if defined(_AIX)
 216     OPENSSL_ppccap_P |= PPC_FPU;
 217
 218     if (sizeof(size_t) == 4) {
 219         struct utsname uts;
 220 # if defined(_SC_AIX_KERNEL_BITMODE)
 221         if (sysconf(_SC_AIX_KERNEL_BITMODE) != 64)
 222             return;
 223 # endif
 224         if (uname(&uts) != 0 || atoi(uts.version) < 6)
 225             return;
 226     }
 227
 228 # if defined(__power_set)
 229     /*
 230      * Value used in __power_set is a single-bit 1<<n one denoting
 231      * specific processor class. Incidentally 0xffffffff<<n can be
 232      * used to denote specific processor and its successors.
 233      */
 234     if (sizeof(size_t) == 4) {
 235         /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 236         if (__power_set(0xffffffffU<<13))       /* POWER5 and later */
 237             OPENSSL_ppccap_P |= PPC_FPU64;
 238     } else {
 239         /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 240         if (__power_set(0x1U<<14))              /* POWER6 */
 241             OPENSSL_ppccap_P |= PPC_FPU64;
 242     }
 243
 244     if (__power_set(0xffffffffU<<14))           /* POWER6 and later */
 245         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 246
 247     if (__power_set(0xffffffffU<<16))           /* POWER8 and later */
 248         OPENSSL_ppccap_P |= PPC_CRYPTO207;
 249
 250     if (__power_set(0xffffffffU<<17))           /* POWER9 and later */
 251         OPENSSL_ppccap_P |= PPC_MADD300;
 252
 253     return;
 254 # endif
 255 #endif
 256
 257     if (getauxval != NULL) {
 258         unsigned long hwcap = getauxval(HWCAP);
 259
 260         if (hwcap & HWCAP_FPU) {
 261             OPENSSL_ppccap_P |= PPC_FPU;
 262
 263             if (sizeof(size_t) == 4) {
 264                 /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 265                 if (hwcap & HWCAP_PPC64)
 266                     OPENSSL_ppccap_P |= PPC_FPU64;
 267             } else {
 268                 /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 269                 if (hwcap & HWCAP_POWER6_EXT)
 270                     OPENSSL_ppccap_P |= PPC_FPU64;
 271             }
 272         }
 273
 274         if (hwcap & HWCAP_ALTIVEC) {
 275             OPENSSL_ppccap_P |= PPC_ALTIVEC;
 276
 277             if ((hwcap & HWCAP_VSX) && (getauxval(HWCAP2) & HWCAP_VEC_CRYPTO))
 278                 OPENSSL_ppccap_P |= PPC_CRYPTO207;
 279         }
 280
 281         if (hwcap & HWCAP_ARCH_3_00) {
 282             OPENSSL_ppccap_P |= PPC_MADD300;
 283         }
 284
 285         return;
 286     }
 287
 288     sigfillset(&all_masked);
 289     sigdelset(&all_masked, SIGILL);
 290     sigdelset(&all_masked, SIGTRAP);
 291 #ifdef SIGEMT
 292     sigdelset(&all_masked, SIGEMT);
 293 #endif
 294     sigdelset(&all_masked, SIGFPE);
 295     sigdelset(&all_masked, SIGBUS);
 296     sigdelset(&all_masked, SIGSEGV);
 297
 298     memset(&ill_act, 0, sizeof(ill_act));
 299     ill_act.sa_handler = ill_handler;
 300     ill_act.sa_mask = all_masked;
 301
 302     sigprocmask(SIG_SETMASK, &ill_act.sa_mask, &oset);
 303     sigaction(SIGILL, &ill_act, &ill_oact);
 304
 305     if (sigsetjmp(ill_jmp,1) == 0) {
 306         OPENSSL_fpu_probe();
 307         OPENSSL_ppccap_P |= PPC_FPU;
 308
 309         if (sizeof(size_t) == 4) {
 310 #ifdef __linux
 311             struct utsname uts;
 312             if (uname(&uts) == 0 && strcmp(uts.machine, "ppc64") == 0)
 313 #endif
 314                 if (sigsetjmp(ill_jmp, 1) == 0) {
 315                     OPENSSL_ppc64_probe();
 316                     OPENSSL_ppccap_P |= PPC_FPU64;
 317                 }
 318         } else {
 319             /*
 320              * Wanted code detecting POWER6 CPU and setting PPC_FPU64
 321              */
 322         }
 323     }
 324
 325     if (sigsetjmp(ill_jmp, 1) == 0) {
 326         OPENSSL_altivec_probe();
 327         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 328         if (sigsetjmp(ill_jmp, 1) == 0) {
 329             OPENSSL_crypto207_probe();
 330             OPENSSL_ppccap_P |= PPC_CRYPTO207;
 331         }
 332     }
 333
 334     if (sigsetjmp(ill_jmp, 1) == 0) {
 335         OPENSSL_madd300_probe();
 336         OPENSSL_ppccap_P |= PPC_MADD300;
 337     }
 338
 339     sigaction(SIGILL, &ill_oact, NULL);
 340     sigprocmask(SIG_SETMASK, &oset, NULL);
 341 }