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 static sigjmp_buf ill_jmp;
  83 static void ill_handler(int sig)
  84 {
  85     siglongjmp(ill_jmp, sig);
  86 }
  87
  88 void OPENSSL_fpu_probe(void);
  89 void OPENSSL_ppc64_probe(void);
  90 void OPENSSL_altivec_probe(void);
  91 void OPENSSL_crypto207_probe(void);
  92
  93 /*
  94  * Use a weak reference to getauxval() so we can use it if it is available
  95  * but don't break the build if it is not. Note that this is *link-time*
  96  * feature detection, not *run-time*. In other words if we link with
  97  * symbol present, it's expected to be present even at run-time.
  98  */
  99 #if defined(__GNUC__) && __GNUC__>=2 && defined(__ELF__)
 100 extern unsigned long getauxval(unsigned long type) __attribute__ ((weak));
 101 #else
 102 static unsigned long (*getauxval) (unsigned long) = NULL;
 103 #endif
 104
 105 /* I wish <sys/auxv.h> was universally available */
 106 #define HWCAP                   16      /* AT_HWCAP */
 107 #define HWCAP_PPC64             (1U << 30)
 108 #define HWCAP_ALTIVEC           (1U << 28)
 109 #define HWCAP_FPU               (1U << 27)
 110 #define HWCAP_POWER6_EXT        (1U << 9)
 111 #define HWCAP_VSX               (1U << 7)
 112
 113 #define HWCAP2                  26      /* AT_HWCAP2 */
 114 #define HWCAP_VEC_CRYPTO        (1U << 25)
 115
 116 # if defined(__GNUC__) && __GNUC__>=2
 117 __attribute__ ((constructor))
 118 # endif
 119 void OPENSSL_cpuid_setup(void)
 120 {
 121     char *e;
 122     struct sigaction ill_oact, ill_act;
 123     sigset_t oset;
 124     static int trigger = 0;
 125
 126     if (trigger)
 127         return;
 128     trigger = 1;
 129
 130     if ((e = getenv("OPENSSL_ppccap"))) {
 131         OPENSSL_ppccap_P = strtoul(e, NULL, 0);
 132         return;
 133     }
 134
 135     OPENSSL_ppccap_P = 0;
 136
 137 #if defined(_AIX)
 138     OPENSSL_ppccap_P |= PPC_FPU;
 139
 140     if (sizeof(size_t) == 4) {
 141         struct utsname uts;
 142 # if defined(_SC_AIX_KERNEL_BITMODE)
 143         if (sysconf(_SC_AIX_KERNEL_BITMODE) != 64)
 144             return;
 145 # endif
 146         if (uname(&uts) != 0 || atoi(uts.version) < 6)
 147             return;
 148     }
 149
 150 # if defined(__power_set)
 151     /*
 152      * Value used in __power_set is a single-bit 1<<n one denoting
 153      * specific processor class. Incidentally 0xffffffff<<n can be
 154      * used to denote specific processor and its successors.
 155      */
 156     if (sizeof(size_t) == 4) {
 157         /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 158         if (__power_set(0xffffffffU<<13))       /* POWER5 and later */
 159             OPENSSL_ppccap_P |= PPC_FPU64;
 160     } else {
 161         /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 162         if (__power_set(0x1U<<14))              /* POWER6 */
 163             OPENSSL_ppccap_P |= PPC_FPU64;
 164     }
 165
 166     if (__power_set(0xffffffffU<<14))           /* POWER6 and later */
 167         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 168
 169     if (__power_set(0xffffffffU<<16))           /* POWER8 and later */
 170         OPENSSL_ppccap_P |= PPC_CRYPTO207;
 171
 172     return;
 173 # endif
 174 #endif
 175
 176     if (getauxval != NULL) {
 177         unsigned long hwcap = getauxval(HWCAP);
 178
 179         if (hwcap & HWCAP_FPU) {
 180             OPENSSL_ppccap_P |= PPC_FPU;
 181
 182             if (sizeof(size_t) == 4) {
 183                 /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 184                 if (hwcap & HWCAP_PPC64)
 185                     OPENSSL_ppccap_P |= PPC_FPU64;
 186             } else {
 187                 /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 188                 if (hwcap & HWCAP_POWER6_EXT)
 189                     OPENSSL_ppccap_P |= PPC_FPU64;
 190             }
 191         }
 192
 193         if (hwcap & HWCAP_ALTIVEC) {
 194             OPENSSL_ppccap_P |= PPC_ALTIVEC;
 195
 196             if ((hwcap & HWCAP_VSX) && (getauxval(HWCAP2) & HWCAP_VEC_CRYPTO))
 197                 OPENSSL_ppccap_P |= PPC_CRYPTO207;
 198         }
 199
 200         return;
 201     }
 202
 203     sigfillset(&all_masked);
 204     sigdelset(&all_masked, SIGILL);
 205     sigdelset(&all_masked, SIGTRAP);
 206 #ifdef SIGEMT
 207     sigdelset(&all_masked, SIGEMT);
 208 #endif
 209     sigdelset(&all_masked, SIGFPE);
 210     sigdelset(&all_masked, SIGBUS);
 211     sigdelset(&all_masked, SIGSEGV);
 212
 213     memset(&ill_act, 0, sizeof(ill_act));
 214     ill_act.sa_handler = ill_handler;
 215     ill_act.sa_mask = all_masked;
 216
 217     sigprocmask(SIG_SETMASK, &ill_act.sa_mask, &oset);
 218     sigaction(SIGILL, &ill_act, &ill_oact);
 219
 220     if (sigsetjmp(ill_jmp,1) == 0) {
 221         OPENSSL_fpu_probe();
 222         OPENSSL_ppccap_P |= PPC_FPU;
 223
 224         if (sizeof(size_t) == 4) {
 225 #ifdef __linux
 226             struct utsname uts;
 227             if (uname(&uts) == 0 && strcmp(uts.machine, "ppc64") == 0)
 228 #endif
 229                 if (sigsetjmp(ill_jmp, 1) == 0) {
 230                     OPENSSL_ppc64_probe();
 231                     OPENSSL_ppccap_P |= PPC_FPU64;
 232                 }
 233         } else {
 234             /*
 235              * Wanted code detecting POWER6 CPU and setting PPC_FPU64
 236              */
 237         }
 238     }
 239
 240     if (sigsetjmp(ill_jmp, 1) == 0) {
 241         OPENSSL_altivec_probe();
 242         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 243         if (sigsetjmp(ill_jmp, 1) == 0) {
 244             OPENSSL_crypto207_probe();
 245             OPENSSL_ppccap_P |= PPC_CRYPTO207;
 246         }
 247     }
 248
 249     sigaction(SIGILL, &ill_oact, NULL);
 250     sigprocmask(SIG_SETMASK, &oset, NULL);
 251 }