crypto/ppccap.c

   1 #include <stdio.h>
   2 #include <stdlib.h>
   3 #include <string.h>
   4 #include <setjmp.h>
   5 #include <signal.h>
   6 #include <unistd.h>
   7 #if defined(__linux) || defined(_AIX)
   8 # include <sys/utsname.h>
   9 #endif
  10 #if defined(_AIX53)     /* defined even on post-5.3 */
  11 # include <sys/systemcfg.h>
  12 # if !defined(__power_set)
  13 #  define __power_set(a) (_system_configuration.implementation & (a))
  14 # endif
  15 #endif
  16 #include <openssl/crypto.h>
  17 #include <openssl/bn.h>
  18
  19 #include "ppc_arch.h"
  20
  21 unsigned int OPENSSL_ppccap_P = 0;
  22
  23 static sigset_t all_masked;
  24
  25 #ifdef OPENSSL_BN_ASM_MONT
  26 int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
  27                 const BN_ULONG *np, const BN_ULONG *n0, int num)
  28 {
  29     int bn_mul_mont_fpu64(BN_ULONG *rp, const BN_ULONG *ap,
  30                           const BN_ULONG *bp, const BN_ULONG *np,
  31                           const BN_ULONG *n0, int num);
  32     int bn_mul_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
  33                         const BN_ULONG *np, const BN_ULONG *n0, int num);
  34
  35     if (sizeof(size_t) == 4) {
  36 # if 1 || (defined(__APPLE__) && defined(__MACH__))
  37         if (num >= 8 && (num & 3) == 0 && (OPENSSL_ppccap_P & PPC_FPU64))
  38             return bn_mul_mont_fpu64(rp, ap, bp, np, n0, num);
  39 # else
  40         /*
  41          * boundary of 32 was experimentally determined on Linux 2.6.22,
  42          * might have to be adjusted on AIX...
  43          */
  44         if (num >= 32 && (num & 3) == 0 && (OPENSSL_ppccap_P & PPC_FPU64)) {
  45             sigset_t oset;
  46             int ret;
  47
  48             sigprocmask(SIG_SETMASK, &all_masked, &oset);
  49             ret = bn_mul_mont_fpu64(rp, ap, bp, np, n0, num);
  50             sigprocmask(SIG_SETMASK, &oset, NULL);
  51
  52             return ret;
  53         }
  54 # endif
  55     } else if ((OPENSSL_ppccap_P & PPC_FPU64))
  56         /*
  57          * this is a "must" on POWER6, but run-time detection is not
  58          * implemented yet...
  59          */
  60         return bn_mul_mont_fpu64(rp, ap, bp, np, n0, num);
  61
  62     return bn_mul_mont_int(rp, ap, bp, np, n0, num);
  63 }
  64 #endif
  65
  66 void sha256_block_p8(void *ctx, const void *inp, size_t len);
  67 void sha256_block_ppc(void *ctx, const void *inp, size_t len);
  68 void sha256_block_data_order(void *ctx, const void *inp, size_t len)
  69 {
  70     OPENSSL_ppccap_P & PPC_CRYPTO207 ? sha256_block_p8(ctx, inp, len) :
  71         sha256_block_ppc(ctx, inp, len);
  72 }
  73
  74 void sha512_block_p8(void *ctx, const void *inp, size_t len);
  75 void sha512_block_ppc(void *ctx, const void *inp, size_t len);
  76 void sha512_block_data_order(void *ctx, const void *inp, size_t len)
  77 {
  78     OPENSSL_ppccap_P & PPC_CRYPTO207 ? sha512_block_p8(ctx, inp, len) :
  79         sha512_block_ppc(ctx, inp, len);
  80 }
  81
  82 void ChaCha20_ctr32_int(unsigned char *out, const unsigned char *inp,
  83                         size_t len, const unsigned int key[8],
  84                         const unsigned int counter[4]);
  85 void ChaCha20_ctr32_vmx(unsigned char *out, const unsigned char *inp,
  86                         size_t len, const unsigned int key[8],
  87                         const unsigned int counter[4]);
  88 void ChaCha20_ctr32(unsigned char *out, const unsigned char *inp,
  89                     size_t len, const unsigned int key[8],
  90                     const unsigned int counter[4])
  91 {
  92     OPENSSL_ppccap_P & PPC_ALTIVEC
  93         ? ChaCha20_ctr32_vmx(out, inp, len, key, counter)
  94         : ChaCha20_ctr32_int(out, inp, len, key, counter);
  95 }
  96
  97 void poly1305_init_int(void *ctx, const unsigned char key[16]);
  98 void poly1305_blocks(void *ctx, const unsigned char *inp, size_t len,
  99                          unsigned int padbit);
 100 void poly1305_emit(void *ctx, unsigned char mac[16],
 101                        const unsigned int nonce[4]);
 102 void poly1305_init_fpu(void *ctx, const unsigned char key[16]);
 103 void poly1305_blocks_fpu(void *ctx, const unsigned char *inp, size_t len,
 104                          unsigned int padbit);
 105 void poly1305_emit_fpu(void *ctx, unsigned char mac[16],
 106                        const unsigned int nonce[4]);
 107 int poly1305_init(void *ctx, const unsigned char key[16], void *func[2])
 108 {
 109     if (sizeof(size_t) == 4 && (OPENSSL_ppccap_P & PPC_FPU)) {
 110         poly1305_init_fpu(ctx,key);
 111         func[0] = poly1305_blocks_fpu;
 112         func[1] = poly1305_emit_fpu;
 113     } else {
 114         poly1305_init_int(ctx,key);
 115         func[0] = poly1305_blocks;
 116         func[1] = poly1305_emit;
 117     }
 118     return 1;
 119 }
 120
 121 static sigjmp_buf ill_jmp;
 122 static void ill_handler(int sig)
 123 {
 124     siglongjmp(ill_jmp, sig);
 125 }
 126
 127 void OPENSSL_fpu_probe(void);
 128 void OPENSSL_ppc64_probe(void);
 129 void OPENSSL_altivec_probe(void);
 130 void OPENSSL_crypto207_probe(void);
 131
 132 /*
 133  * Use a weak reference to getauxval() so we can use it if it is available
 134  * but don't break the build if it is not. Note that this is *link-time*
 135  * feature detection, not *run-time*. In other words if we link with
 136  * symbol present, it's expected to be present even at run-time.
 137  */
 138 #if defined(__GNUC__) && __GNUC__>=2 && defined(__ELF__)
 139 extern unsigned long getauxval(unsigned long type) __attribute__ ((weak));
 140 #else
 141 static unsigned long (*getauxval) (unsigned long) = NULL;
 142 #endif
 143
 144 /* I wish <sys/auxv.h> was universally available */
 145 #define HWCAP                   16      /* AT_HWCAP */
 146 #define HWCAP_PPC64             (1U << 30)
 147 #define HWCAP_ALTIVEC           (1U << 28)
 148 #define HWCAP_FPU               (1U << 27)
 149 #define HWCAP_POWER6_EXT        (1U << 9)
 150 #define HWCAP_VSX               (1U << 7)
 151
 152 #define HWCAP2                  26      /* AT_HWCAP2 */
 153 #define HWCAP_VEC_CRYPTO        (1U << 25)
 154
 155 # if defined(__GNUC__) && __GNUC__>=2
 156 __attribute__ ((constructor))
 157 # endif
 158 void OPENSSL_cpuid_setup(void)
 159 {
 160     char *e;
 161     struct sigaction ill_oact, ill_act;
 162     sigset_t oset;
 163     static int trigger = 0;
 164
 165     if (trigger)
 166         return;
 167     trigger = 1;
 168
 169     if ((e = getenv("OPENSSL_ppccap"))) {
 170         OPENSSL_ppccap_P = strtoul(e, NULL, 0);
 171         return;
 172     }
 173
 174     OPENSSL_ppccap_P = 0;
 175
 176 #if defined(_AIX)
 177     OPENSSL_ppccap_P |= PPC_FPU;
 178
 179     if (sizeof(size_t) == 4) {
 180         struct utsname uts;
 181 # if defined(_SC_AIX_KERNEL_BITMODE)
 182         if (sysconf(_SC_AIX_KERNEL_BITMODE) != 64)
 183             return;
 184 # endif
 185         if (uname(&uts) != 0 || atoi(uts.version) < 6)
 186             return;
 187     }
 188
 189 # if defined(__power_set)
 190     /*
 191      * Value used in __power_set is a single-bit 1<<n one denoting
 192      * specific processor class. Incidentally 0xffffffff<<n can be
 193      * used to denote specific processor and its successors.
 194      */
 195     if (sizeof(size_t) == 4) {
 196         /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 197         if (__power_set(0xffffffffU<<13))       /* POWER5 and later */
 198             OPENSSL_ppccap_P |= PPC_FPU64;
 199     } else {
 200         /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 201         if (__power_set(0x1U<<14))              /* POWER6 */
 202             OPENSSL_ppccap_P |= PPC_FPU64;
 203     }
 204
 205     if (__power_set(0xffffffffU<<14))           /* POWER6 and later */
 206         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 207
 208     if (__power_set(0xffffffffU<<16))           /* POWER8 and later */
 209         OPENSSL_ppccap_P |= PPC_CRYPTO207;
 210
 211     return;
 212 # endif
 213 #endif
 214
 215     if (getauxval != NULL) {
 216         unsigned long hwcap = getauxval(HWCAP);
 217
 218         if (hwcap & HWCAP_FPU) {
 219             OPENSSL_ppccap_P |= PPC_FPU;
 220
 221             if (sizeof(size_t) == 4) {
 222                 /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 223                 if (hwcap & HWCAP_PPC64)
 224                     OPENSSL_ppccap_P |= PPC_FPU64;
 225             } else {
 226                 /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 227                 if (hwcap & HWCAP_POWER6_EXT)
 228                     OPENSSL_ppccap_P |= PPC_FPU64;
 229             }
 230         }
 231
 232         if (hwcap & HWCAP_ALTIVEC) {
 233             OPENSSL_ppccap_P |= PPC_ALTIVEC;
 234
 235             if ((hwcap & HWCAP_VSX) && (getauxval(HWCAP2) & HWCAP_VEC_CRYPTO))
 236                 OPENSSL_ppccap_P |= PPC_CRYPTO207;
 237         }
 238
 239         return;
 240     }
 241
 242     sigfillset(&all_masked);
 243     sigdelset(&all_masked, SIGILL);
 244     sigdelset(&all_masked, SIGTRAP);
 245 #ifdef SIGEMT
 246     sigdelset(&all_masked, SIGEMT);
 247 #endif
 248     sigdelset(&all_masked, SIGFPE);
 249     sigdelset(&all_masked, SIGBUS);
 250     sigdelset(&all_masked, SIGSEGV);
 251
 252     memset(&ill_act, 0, sizeof(ill_act));
 253     ill_act.sa_handler = ill_handler;
 254     ill_act.sa_mask = all_masked;
 255
 256     sigprocmask(SIG_SETMASK, &ill_act.sa_mask, &oset);
 257     sigaction(SIGILL, &ill_act, &ill_oact);
 258
 259     if (sigsetjmp(ill_jmp,1) == 0) {
 260         OPENSSL_fpu_probe();
 261         OPENSSL_ppccap_P |= PPC_FPU;
 262
 263         if (sizeof(size_t) == 4) {
 264 #ifdef __linux
 265             struct utsname uts;
 266             if (uname(&uts) == 0 && strcmp(uts.machine, "ppc64") == 0)
 267 #endif
 268                 if (sigsetjmp(ill_jmp, 1) == 0) {
 269                     OPENSSL_ppc64_probe();
 270                     OPENSSL_ppccap_P |= PPC_FPU64;
 271                 }
 272         } else {
 273             /*
 274              * Wanted code detecting POWER6 CPU and setting PPC_FPU64
 275              */
 276         }
 277     }
 278
 279     if (sigsetjmp(ill_jmp, 1) == 0) {
 280         OPENSSL_altivec_probe();
 281         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 282         if (sigsetjmp(ill_jmp, 1) == 0) {
 283             OPENSSL_crypto207_probe();
 284             OPENSSL_ppccap_P |= PPC_CRYPTO207;
 285         }
 286     }
 287
 288     sigaction(SIGILL, &ill_oact, NULL);
 289     sigprocmask(SIG_SETMASK, &oset, NULL);
 290 }