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_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
  39                         const BN_ULONG *np, const BN_ULONG *n0, int num);
  40     int bn_mul4x_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
  41                           const BN_ULONG *np, const BN_ULONG *n0, int num);
  42
  43     if (num < 4)
  44         return 0;
  45
  46     if ((num & 3) == 0)
  47         return bn_mul4x_mont_int(rp, ap, bp, np, n0, num);
  48
  49     /*
  50      * There used to be [optional] call to bn_mul_mont_fpu64 here,
  51      * but above subroutine is faster on contemporary processors.
  52      * Formulation means that there might be old processors where
  53      * FPU code path would be faster, POWER6 perhaps, but there was
  54      * no opportunity to figure it out...
  55      */
  56
  57     return bn_mul_mont_int(rp, ap, bp, np, n0, num);
  58 }
  59 #endif
  60
  61 void sha256_block_p8(void *ctx, const void *inp, size_t len);
  62 void sha256_block_ppc(void *ctx, const void *inp, size_t len);
  63 void sha256_block_data_order(void *ctx, const void *inp, size_t len)
  64 {
  65     OPENSSL_ppccap_P & PPC_CRYPTO207 ? sha256_block_p8(ctx, inp, len) :
  66         sha256_block_ppc(ctx, inp, len);
  67 }
  68
  69 void sha512_block_p8(void *ctx, const void *inp, size_t len);
  70 void sha512_block_ppc(void *ctx, const void *inp, size_t len);
  71 void sha512_block_data_order(void *ctx, const void *inp, size_t len)
  72 {
  73     OPENSSL_ppccap_P & PPC_CRYPTO207 ? sha512_block_p8(ctx, inp, len) :
  74         sha512_block_ppc(ctx, inp, len);
  75 }
  76
  77 #ifndef OPENSSL_NO_CHACHA
  78 void ChaCha20_ctr32_int(unsigned char *out, const unsigned char *inp,
  79                         size_t len, const unsigned int key[8],
  80                         const unsigned int counter[4]);
  81 void ChaCha20_ctr32_vmx(unsigned char *out, const unsigned char *inp,
  82                         size_t len, const unsigned int key[8],
  83                         const unsigned int counter[4]);
  84 void ChaCha20_ctr32(unsigned char *out, const unsigned char *inp,
  85                     size_t len, const unsigned int key[8],
  86                     const unsigned int counter[4])
  87 {
  88     OPENSSL_ppccap_P & PPC_ALTIVEC
  89         ? ChaCha20_ctr32_vmx(out, inp, len, key, counter)
  90         : ChaCha20_ctr32_int(out, inp, len, key, counter);
  91 }
  92 #endif
  93
  94 #ifndef OPENSSL_NO_POLY1305
  95 void poly1305_init_int(void *ctx, const unsigned char key[16]);
  96 void poly1305_blocks(void *ctx, const unsigned char *inp, size_t len,
  97                          unsigned int padbit);
  98 void poly1305_emit(void *ctx, unsigned char mac[16],
  99                        const unsigned int nonce[4]);
 100 void poly1305_init_fpu(void *ctx, const unsigned char key[16]);
 101 void poly1305_blocks_fpu(void *ctx, const unsigned char *inp, size_t len,
 102                          unsigned int padbit);
 103 void poly1305_emit_fpu(void *ctx, unsigned char mac[16],
 104                        const unsigned int nonce[4]);
 105 int poly1305_init(void *ctx, const unsigned char key[16], void *func[2])
 106 {
 107     if (sizeof(size_t) == 4 && (OPENSSL_ppccap_P & PPC_FPU)) {
 108         poly1305_init_fpu(ctx, key);
 109         func[0] = poly1305_blocks_fpu;
 110         func[1] = poly1305_emit_fpu;
 111     } else {
 112         poly1305_init_int(ctx, key);
 113         func[0] = poly1305_blocks;
 114         func[1] = poly1305_emit;
 115     }
 116     return 1;
 117 }
 118 #endif
 119
 120 #ifdef ECP_NISTZ256_ASM
 121 void ecp_nistz256_mul_mont(unsigned long res[4], const unsigned long a[4],
 122                            const unsigned long b[4]);
 123
 124 void ecp_nistz256_to_mont(unsigned long res[4], const unsigned long in[4]);
 125 void ecp_nistz256_to_mont(unsigned long res[4], const unsigned long in[4])
 126 {
 127     static const unsigned long RR[] = { 0x0000000000000003U,
 128                                         0xfffffffbffffffffU,
 129                                         0xfffffffffffffffeU,
 130                                         0x00000004fffffffdU };
 131
 132     ecp_nistz256_mul_mont(res, in, RR);
 133 }
 134
 135 void ecp_nistz256_from_mont(unsigned long res[4], const unsigned long in[4]);
 136 void ecp_nistz256_from_mont(unsigned long res[4], const unsigned long in[4])
 137 {
 138     static const unsigned long one[] = { 1, 0, 0, 0 };
 139
 140     ecp_nistz256_mul_mont(res, in, one);
 141 }
 142 #endif
 143
 144 static sigjmp_buf ill_jmp;
 145 static void ill_handler(int sig)
 146 {
 147     siglongjmp(ill_jmp, sig);
 148 }
 149
 150 void OPENSSL_fpu_probe(void);
 151 void OPENSSL_ppc64_probe(void);
 152 void OPENSSL_altivec_probe(void);
 153 void OPENSSL_crypto207_probe(void);
 154 void OPENSSL_madd300_probe(void);
 155
 156 /*
 157  * Use a weak reference to getauxval() so we can use it if it is available
 158  * but don't break the build if it is not. Note that this is *link-time*
 159  * feature detection, not *run-time*. In other words if we link with
 160  * symbol present, it's expected to be present even at run-time.
 161  */
 162 #if defined(__GNUC__) && __GNUC__>=2 && defined(__ELF__)
 163 extern unsigned long getauxval(unsigned long type) __attribute__ ((weak));
 164 #else
 165 static unsigned long (*getauxval) (unsigned long) = NULL;
 166 #endif
 167
 168 /* I wish <sys/auxv.h> was universally available */
 169 #define HWCAP                   16      /* AT_HWCAP */
 170 #define HWCAP_PPC64             (1U << 30)
 171 #define HWCAP_ALTIVEC           (1U << 28)
 172 #define HWCAP_FPU               (1U << 27)
 173 #define HWCAP_POWER6_EXT        (1U << 9)
 174 #define HWCAP_VSX               (1U << 7)
 175
 176 #define HWCAP2                  26      /* AT_HWCAP2 */
 177 #define HWCAP_VEC_CRYPTO        (1U << 25)
 178 #define HWCAP_ARCH_3_00         (1U << 23)
 179
 180 # if defined(__GNUC__) && __GNUC__>=2
 181 __attribute__ ((constructor))
 182 # endif
 183 void OPENSSL_cpuid_setup(void)
 184 {
 185     char *e;
 186     struct sigaction ill_oact, ill_act;
 187     sigset_t oset;
 188     static int trigger = 0;
 189
 190     if (trigger)
 191         return;
 192     trigger = 1;
 193
 194     if ((e = getenv("OPENSSL_ppccap"))) {
 195         OPENSSL_ppccap_P = strtoul(e, NULL, 0);
 196         return;
 197     }
 198
 199     OPENSSL_ppccap_P = 0;
 200
 201 #if defined(_AIX)
 202     OPENSSL_ppccap_P |= PPC_FPU;
 203
 204     if (sizeof(size_t) == 4) {
 205         struct utsname uts;
 206 # if defined(_SC_AIX_KERNEL_BITMODE)
 207         if (sysconf(_SC_AIX_KERNEL_BITMODE) != 64)
 208             return;
 209 # endif
 210         if (uname(&uts) != 0 || atoi(uts.version) < 6)
 211             return;
 212     }
 213
 214 # if defined(__power_set)
 215     /*
 216      * Value used in __power_set is a single-bit 1<<n one denoting
 217      * specific processor class. Incidentally 0xffffffff<<n can be
 218      * used to denote specific processor and its successors.
 219      */
 220     if (sizeof(size_t) == 4) {
 221         /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 222         if (__power_set(0xffffffffU<<13))       /* POWER5 and later */
 223             OPENSSL_ppccap_P |= PPC_FPU64;
 224     } else {
 225         /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 226         if (__power_set(0x1U<<14))              /* POWER6 */
 227             OPENSSL_ppccap_P |= PPC_FPU64;
 228     }
 229
 230     if (__power_set(0xffffffffU<<14))           /* POWER6 and later */
 231         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 232
 233     if (__power_set(0xffffffffU<<16))           /* POWER8 and later */
 234         OPENSSL_ppccap_P |= PPC_CRYPTO207;
 235
 236     if (__power_set(0xffffffffU<<17))           /* POWER9 and later */
 237         OPENSSL_ppccap_P |= PPC_MADD300;
 238
 239     return;
 240 # endif
 241 #endif
 242
 243     if (getauxval != NULL) {
 244         unsigned long hwcap = getauxval(HWCAP);
 245
 246         if (hwcap & HWCAP_FPU) {
 247             OPENSSL_ppccap_P |= PPC_FPU;
 248
 249             if (sizeof(size_t) == 4) {
 250                 /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 251                 if (hwcap & HWCAP_PPC64)
 252                     OPENSSL_ppccap_P |= PPC_FPU64;
 253             } else {
 254                 /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 255                 if (hwcap & HWCAP_POWER6_EXT)
 256                     OPENSSL_ppccap_P |= PPC_FPU64;
 257             }
 258         }
 259
 260         if (hwcap & HWCAP_ALTIVEC) {
 261             OPENSSL_ppccap_P |= PPC_ALTIVEC;
 262
 263             if ((hwcap & HWCAP_VSX) && (getauxval(HWCAP2) & HWCAP_VEC_CRYPTO))
 264                 OPENSSL_ppccap_P |= PPC_CRYPTO207;
 265         }
 266
 267         if (hwcap & HWCAP_ARCH_3_00) {
 268             OPENSSL_ppccap_P |= PPC_MADD300;
 269         }
 270
 271         return;
 272     }
 273
 274     sigfillset(&all_masked);
 275     sigdelset(&all_masked, SIGILL);
 276     sigdelset(&all_masked, SIGTRAP);
 277 #ifdef SIGEMT
 278     sigdelset(&all_masked, SIGEMT);
 279 #endif
 280     sigdelset(&all_masked, SIGFPE);
 281     sigdelset(&all_masked, SIGBUS);
 282     sigdelset(&all_masked, SIGSEGV);
 283
 284     memset(&ill_act, 0, sizeof(ill_act));
 285     ill_act.sa_handler = ill_handler;
 286     ill_act.sa_mask = all_masked;
 287
 288     sigprocmask(SIG_SETMASK, &ill_act.sa_mask, &oset);
 289     sigaction(SIGILL, &ill_act, &ill_oact);
 290
 291     if (sigsetjmp(ill_jmp,1) == 0) {
 292         OPENSSL_fpu_probe();
 293         OPENSSL_ppccap_P |= PPC_FPU;
 294
 295         if (sizeof(size_t) == 4) {
 296 #ifdef __linux
 297             struct utsname uts;
 298             if (uname(&uts) == 0 && strcmp(uts.machine, "ppc64") == 0)
 299 #endif
 300                 if (sigsetjmp(ill_jmp, 1) == 0) {
 301                     OPENSSL_ppc64_probe();
 302                     OPENSSL_ppccap_P |= PPC_FPU64;
 303                 }
 304         } else {
 305             /*
 306              * Wanted code detecting POWER6 CPU and setting PPC_FPU64
 307              */
 308         }
 309     }
 310
 311     if (sigsetjmp(ill_jmp, 1) == 0) {
 312         OPENSSL_altivec_probe();
 313         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 314         if (sigsetjmp(ill_jmp, 1) == 0) {
 315             OPENSSL_crypto207_probe();
 316             OPENSSL_ppccap_P |= PPC_CRYPTO207;
 317         }
 318     }
 319
 320     if (sigsetjmp(ill_jmp, 1) == 0) {
 321         OPENSSL_madd300_probe();
 322         OPENSSL_ppccap_P |= PPC_MADD300;
 323     }
 324
 325     sigaction(SIGILL, &ill_oact, NULL);
 326     sigprocmask(SIG_SETMASK, &oset, NULL);
 327 }