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 #ifndef OPENSSL_NO_CHACHA
  83 void ChaCha20_ctr32_int(unsigned char *out, const unsigned char *inp,
  84                         size_t len, const unsigned int key[8],
  85                         const unsigned int counter[4]);
  86 void ChaCha20_ctr32_vmx(unsigned char *out, const unsigned char *inp,
  87                         size_t len, const unsigned int key[8],
  88                         const unsigned int counter[4]);
  89 void ChaCha20_ctr32(unsigned char *out, const unsigned char *inp,
  90                     size_t len, const unsigned int key[8],
  91                     const unsigned int counter[4])
  92 {
  93     OPENSSL_ppccap_P & PPC_ALTIVEC
  94         ? ChaCha20_ctr32_vmx(out, inp, len, key, counter)
  95         : ChaCha20_ctr32_int(out, inp, len, key, counter);
  96 }
  97 #endif
  98
  99 #ifndef OPENSSL_NO_POLY1305
 100 void poly1305_init_int(void *ctx, const unsigned char key[16]);
 101 void poly1305_blocks(void *ctx, const unsigned char *inp, size_t len,
 102                          unsigned int padbit);
 103 void poly1305_emit(void *ctx, unsigned char mac[16],
 104                        const unsigned int nonce[4]);
 105 void poly1305_init_fpu(void *ctx, const unsigned char key[16]);
 106 void poly1305_blocks_fpu(void *ctx, const unsigned char *inp, size_t len,
 107                          unsigned int padbit);
 108 void poly1305_emit_fpu(void *ctx, unsigned char mac[16],
 109                        const unsigned int nonce[4]);
 110 int poly1305_init(void *ctx, const unsigned char key[16], void *func[2])
 111 {
 112     if (sizeof(size_t) == 4 && (OPENSSL_ppccap_P & PPC_FPU)) {
 113         poly1305_init_fpu(ctx,key);
 114         func[0] = poly1305_blocks_fpu;
 115         func[1] = poly1305_emit_fpu;
 116     } else {
 117         poly1305_init_int(ctx,key);
 118         func[0] = poly1305_blocks;
 119         func[1] = poly1305_emit;
 120     }
 121     return 1;
 122 }
 123 #endif
 124
 125 static sigjmp_buf ill_jmp;
 126 static void ill_handler(int sig)
 127 {
 128     siglongjmp(ill_jmp, sig);
 129 }
 130
 131 void OPENSSL_fpu_probe(void);
 132 void OPENSSL_ppc64_probe(void);
 133 void OPENSSL_altivec_probe(void);
 134 void OPENSSL_crypto207_probe(void);
 135
 136 /*
 137  * Use a weak reference to getauxval() so we can use it if it is available
 138  * but don't break the build if it is not. Note that this is *link-time*
 139  * feature detection, not *run-time*. In other words if we link with
 140  * symbol present, it's expected to be present even at run-time.
 141  */
 142 #if defined(__GNUC__) && __GNUC__>=2 && defined(__ELF__)
 143 extern unsigned long getauxval(unsigned long type) __attribute__ ((weak));
 144 #else
 145 static unsigned long (*getauxval) (unsigned long) = NULL;
 146 #endif
 147
 148 /* I wish <sys/auxv.h> was universally available */
 149 #define HWCAP                   16      /* AT_HWCAP */
 150 #define HWCAP_PPC64             (1U << 30)
 151 #define HWCAP_ALTIVEC           (1U << 28)
 152 #define HWCAP_FPU               (1U << 27)
 153 #define HWCAP_POWER6_EXT        (1U << 9)
 154 #define HWCAP_VSX               (1U << 7)
 155
 156 #define HWCAP2                  26      /* AT_HWCAP2 */
 157 #define HWCAP_VEC_CRYPTO        (1U << 25)
 158 #define HWCAP_ARCH_3_00         (1U << 23)
 159
 160 # if defined(__GNUC__) && __GNUC__>=2
 161 __attribute__ ((constructor))
 162 # endif
 163 void OPENSSL_cpuid_setup(void)
 164 {
 165     char *e;
 166     struct sigaction ill_oact, ill_act;
 167     sigset_t oset;
 168     static int trigger = 0;
 169
 170     if (trigger)
 171         return;
 172     trigger = 1;
 173
 174     if ((e = getenv("OPENSSL_ppccap"))) {
 175         OPENSSL_ppccap_P = strtoul(e, NULL, 0);
 176         return;
 177     }
 178
 179     OPENSSL_ppccap_P = 0;
 180
 181 #if defined(_AIX)
 182     OPENSSL_ppccap_P |= PPC_FPU;
 183
 184     if (sizeof(size_t) == 4) {
 185         struct utsname uts;
 186 # if defined(_SC_AIX_KERNEL_BITMODE)
 187         if (sysconf(_SC_AIX_KERNEL_BITMODE) != 64)
 188             return;
 189 # endif
 190         if (uname(&uts) != 0 || atoi(uts.version) < 6)
 191             return;
 192     }
 193
 194 # if defined(__power_set)
 195     /*
 196      * Value used in __power_set is a single-bit 1<<n one denoting
 197      * specific processor class. Incidentally 0xffffffff<<n can be
 198      * used to denote specific processor and its successors.
 199      */
 200     if (sizeof(size_t) == 4) {
 201         /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 202         if (__power_set(0xffffffffU<<13))       /* POWER5 and later */
 203             OPENSSL_ppccap_P |= PPC_FPU64;
 204     } else {
 205         /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 206         if (__power_set(0x1U<<14))              /* POWER6 */
 207             OPENSSL_ppccap_P |= PPC_FPU64;
 208     }
 209
 210     if (__power_set(0xffffffffU<<14))           /* POWER6 and later */
 211         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 212
 213     if (__power_set(0xffffffffU<<16))           /* POWER8 and later */
 214         OPENSSL_ppccap_P |= PPC_CRYPTO207;
 215
 216     if (__power_set(0xffffffffU<<17))           /* POWER9 and later */
 217         OPENSSL_ppccap_P |= PPC_MADD300;
 218
 219     return;
 220 # endif
 221 #endif
 222
 223     if (getauxval != NULL) {
 224         unsigned long hwcap = getauxval(HWCAP);
 225
 226         if (hwcap & HWCAP_FPU) {
 227             OPENSSL_ppccap_P |= PPC_FPU;
 228
 229             if (sizeof(size_t) == 4) {
 230                 /* In 32-bit case PPC_FPU64 is always fastest [if option] */
 231                 if (hwcap & HWCAP_PPC64)
 232                     OPENSSL_ppccap_P |= PPC_FPU64;
 233             } else {
 234                 /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */
 235                 if (hwcap & HWCAP_POWER6_EXT)
 236                     OPENSSL_ppccap_P |= PPC_FPU64;
 237             }
 238         }
 239
 240         if (hwcap & HWCAP_ALTIVEC) {
 241             OPENSSL_ppccap_P |= PPC_ALTIVEC;
 242
 243             if ((hwcap & HWCAP_VSX) && (getauxval(HWCAP2) & HWCAP_VEC_CRYPTO))
 244                 OPENSSL_ppccap_P |= PPC_CRYPTO207;
 245         }
 246
 247         if (hwcap & HWCAP_ARCH_3_00) {
 248             OPENSSL_ppccap_P |= PPC_MADD300;
 249         }
 250
 251         return;
 252     }
 253
 254     sigfillset(&all_masked);
 255     sigdelset(&all_masked, SIGILL);
 256     sigdelset(&all_masked, SIGTRAP);
 257 #ifdef SIGEMT
 258     sigdelset(&all_masked, SIGEMT);
 259 #endif
 260     sigdelset(&all_masked, SIGFPE);
 261     sigdelset(&all_masked, SIGBUS);
 262     sigdelset(&all_masked, SIGSEGV);
 263
 264     memset(&ill_act, 0, sizeof(ill_act));
 265     ill_act.sa_handler = ill_handler;
 266     ill_act.sa_mask = all_masked;
 267
 268     sigprocmask(SIG_SETMASK, &ill_act.sa_mask, &oset);
 269     sigaction(SIGILL, &ill_act, &ill_oact);
 270
 271     if (sigsetjmp(ill_jmp,1) == 0) {
 272         OPENSSL_fpu_probe();
 273         OPENSSL_ppccap_P |= PPC_FPU;
 274
 275         if (sizeof(size_t) == 4) {
 276 #ifdef __linux
 277             struct utsname uts;
 278             if (uname(&uts) == 0 && strcmp(uts.machine, "ppc64") == 0)
 279 #endif
 280                 if (sigsetjmp(ill_jmp, 1) == 0) {
 281                     OPENSSL_ppc64_probe();
 282                     OPENSSL_ppccap_P |= PPC_FPU64;
 283                 }
 284         } else {
 285             /*
 286              * Wanted code detecting POWER6 CPU and setting PPC_FPU64
 287              */
 288         }
 289     }
 290
 291     if (sigsetjmp(ill_jmp, 1) == 0) {
 292         OPENSSL_altivec_probe();
 293         OPENSSL_ppccap_P |= PPC_ALTIVEC;
 294         if (sigsetjmp(ill_jmp, 1) == 0) {
 295             OPENSSL_crypto207_probe();
 296             OPENSSL_ppccap_P |= PPC_CRYPTO207;
 297         }
 298     }
 299
 300     if (sigsetjmp(ill_jmp, 1) == 0) {
 301         OPENSSL_madd300_probe();
 302         OPENSSL_ppccap_P |= PPC_MADD300;
 303     }
 304
 305     sigaction(SIGILL, &ill_oact, NULL);
 306     sigprocmask(SIG_SETMASK, &oset, NULL);
 307 }