1 #include <stdio.h> 2 #include <stdlib.h> 3 #include <string.h> 4 #include <setjmp.h> 5 #include <signal.h> 6 #include <sys/time.h> 7 #include <unistd.h> 8 #include <openssl/bn.h> 9 10 #include "sparc_arch.h" 11 12 #if defined(__GNUC__) && defined(__linux) 13 __attribute__ ((visibility("hidden"))) 14 #endif 15 unsigned int OPENSSL_sparcv9cap_P[2] = { SPARCV9_TICK_PRIVILEGED, 0 }; 16 17 int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, 18 const BN_ULONG *np, const BN_ULONG *n0, int num) 19 { 20 int bn_mul_mont_vis3(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, 21 const BN_ULONG *np, const BN_ULONG *n0, int num); 22 int bn_mul_mont_fpu(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, 23 const BN_ULONG *np, const BN_ULONG *n0, int num); 24 int bn_mul_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, 25 const BN_ULONG *np, const BN_ULONG *n0, int num); 26 27 if (!(num & 1) && num >= 6) { 28 if ((num & 15) == 0 && num <= 64 && 29 (OPENSSL_sparcv9cap_P[1] & (CFR_MONTMUL | CFR_MONTSQR)) == 30 (CFR_MONTMUL | CFR_MONTSQR)) { 31 typedef int (*bn_mul_mont_f) (BN_ULONG *rp, const BN_ULONG *ap, 32 const BN_ULONG *bp, 33 const BN_ULONG *np, 34 const BN_ULONG *n0); 35 int bn_mul_mont_t4_8(BN_ULONG *rp, const BN_ULONG *ap, 36 const BN_ULONG *bp, const BN_ULONG *np, 37 const BN_ULONG *n0); 38 int bn_mul_mont_t4_16(BN_ULONG *rp, const BN_ULONG *ap, 39 const BN_ULONG *bp, const BN_ULONG *np, 40 const BN_ULONG *n0); 41 int bn_mul_mont_t4_24(BN_ULONG *rp, const BN_ULONG *ap, 42 const BN_ULONG *bp, const BN_ULONG *np, 43 const BN_ULONG *n0); 44 int bn_mul_mont_t4_32(BN_ULONG *rp, const BN_ULONG *ap, 45 const BN_ULONG *bp, const BN_ULONG *np, 46 const BN_ULONG *n0); 47 static const bn_mul_mont_f funcs[4] = { 48 bn_mul_mont_t4_8, bn_mul_mont_t4_16, 49 bn_mul_mont_t4_24, bn_mul_mont_t4_32 50 }; 51 bn_mul_mont_f worker = funcs[num / 16 - 1]; 52 53 if ((*worker) (rp, ap, bp, np, n0)) 54 return 1; 55 /* retry once and fall back */ 56 if ((*worker) (rp, ap, bp, np, n0)) 57 return 1; 58 return bn_mul_mont_vis3(rp, ap, bp, np, n0, num); 59 } 60 if ((OPENSSL_sparcv9cap_P[0] & SPARCV9_VIS3)) 61 return bn_mul_mont_vis3(rp, ap, bp, np, n0, num); 62 else if (num >= 8 && 63 (OPENSSL_sparcv9cap_P[0] & 64 (SPARCV9_PREFER_FPU | SPARCV9_VIS1)) == 65 (SPARCV9_PREFER_FPU | SPARCV9_VIS1)) 66 return bn_mul_mont_fpu(rp, ap, bp, np, n0, num); 67 } 68 return bn_mul_mont_int(rp, ap, bp, np, n0, num); 69 } 70 71 unsigned long _sparcv9_rdtick(void); 72 void _sparcv9_vis1_probe(void); 73 unsigned long _sparcv9_vis1_instrument(void); 74 void _sparcv9_vis2_probe(void); 75 void _sparcv9_fmadd_probe(void); 76 unsigned long _sparcv9_rdcfr(void); 77 void _sparcv9_vis3_probe(void); 78 unsigned long _sparcv9_random(void); 79 size_t _sparcv9_vis1_instrument_bus(unsigned int *, size_t); 80 size_t _sparcv9_vis1_instrument_bus2(unsigned int *, size_t, size_t); 81 82 unsigned long OPENSSL_rdtsc(void) 83 { 84 if (OPENSSL_sparcv9cap_P[0] & SPARCV9_TICK_PRIVILEGED) 85 #if defined(__sun) && defined(__SVR4) 86 return gethrtime(); 87 #else 88 return 0; 89 #endif 90 else 91 return _sparcv9_rdtick(); 92 } 93 94 size_t OPENSSL_instrument_bus(unsigned int *out, size_t cnt) 95 { 96 if ((OPENSSL_sparcv9cap_P[0] & (SPARCV9_TICK_PRIVILEGED | SPARCV9_BLK)) == 97 SPARCV9_BLK) 98 return _sparcv9_vis1_instrument_bus(out, cnt); 99 else 100 return 0; 101 } 102 103 size_t OPENSSL_instrument_bus2(unsigned int *out, size_t cnt, size_t max) 104 { 105 if ((OPENSSL_sparcv9cap_P[0] & (SPARCV9_TICK_PRIVILEGED | SPARCV9_BLK)) == 106 SPARCV9_BLK) 107 return _sparcv9_vis1_instrument_bus2(out, cnt, max); 108 else 109 return 0; 110 } 111 112 #if 0 && defined(__sun) && defined(__SVR4) 113 /* 114 * This code path is disabled, because of incompatibility of libdevinfo.so.1 115 * and libmalloc.so.1 (see below for details) 116 */ 117 # include <malloc.h> 118 # include <dlfcn.h> 119 # include <libdevinfo.h> 120 # include <sys/systeminfo.h> 121 122 typedef di_node_t(*di_init_t) (const char *, uint_t); 123 typedef void (*di_fini_t) (di_node_t); 124 typedef char *(*di_node_name_t) (di_node_t); 125 typedef int (*di_walk_node_t) (di_node_t, uint_t, di_node_name_t, 126 int (*)(di_node_t, di_node_name_t)); 127 128 # define DLLINK(h,name) (name=(name##_t)dlsym((h),#name)) 129 130 static int walk_nodename(di_node_t node, di_node_name_t di_node_name) 131 { 132 char *name = (*di_node_name) (node); 133 134 /* This is expected to catch all UltraSPARC flavors prior T1 */ 135 if (!strcmp(name, "SUNW,UltraSPARC") || 136 /* covers II,III,IV */ 137 !strncmp(name, "SUNW,UltraSPARC-I", 17)) { 138 OPENSSL_sparcv9cap_P[0] |= SPARCV9_PREFER_FPU | SPARCV9_VIS1; 139 140 /* %tick is privileged only on UltraSPARC-I/II, but not IIe */ 141 if (name[14] != '\0' && name[17] != '\0' && name[18] != '\0') 142 OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED; 143 144 return DI_WALK_TERMINATE; 145 } 146 /* This is expected to catch remaining UltraSPARCs, such as T1 */ 147 else if (!strncmp(name, "SUNW,UltraSPARC", 15)) { 148 OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED; 149 150 return DI_WALK_TERMINATE; 151 } 152 153 return DI_WALK_CONTINUE; 154 } 155 156 void OPENSSL_cpuid_setup(void) 157 { 158 void *h; 159 char *e, si[256]; 160 static int trigger = 0; 161 162 if (trigger) 163 return; 164 trigger = 1; 165 166 if ((e = getenv("OPENSSL_sparcv9cap"))) { 167 OPENSSL_sparcv9cap_P[0] = strtoul(e, NULL, 0); 168 return; 169 } 170 171 if (sysinfo(SI_MACHINE, si, sizeof(si)) > 0) { 172 if (strcmp(si, "sun4v")) 173 /* FPU is preferred for all CPUs, but US-T1/2 */ 174 OPENSSL_sparcv9cap_P[0] |= SPARCV9_PREFER_FPU; 175 } 176 177 if (sysinfo(SI_ISALIST, si, sizeof(si)) > 0) { 178 if (strstr(si, "+vis")) 179 OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS1 | SPARCV9_BLK; 180 if (strstr(si, "+vis2")) { 181 OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS2; 182 OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED; 183 return; 184 } 185 } 186 # ifdef M_KEEP 187 /* 188 * Solaris libdevinfo.so.1 is effectively incomatible with 189 * libmalloc.so.1. Specifically, if application is linked with 190 * -lmalloc, it crashes upon startup with SIGSEGV in 191 * free(3LIBMALLOC) called by di_fini. Prior call to 192 * mallopt(M_KEEP,0) somehow helps... But not always... 193 */ 194 if ((h = dlopen(NULL, RTLD_LAZY))) { 195 union { 196 void *p; 197 int (*f) (int, int); 198 } sym; 199 if ((sym.p = dlsym(h, "mallopt"))) 200 (*sym.f) (M_KEEP, 0); 201 dlclose(h); 202 } 203 # endif 204 if ((h = dlopen("libdevinfo.so.1", RTLD_LAZY))) 205 do { 206 di_init_t di_init; 207 di_fini_t di_fini; 208 di_walk_node_t di_walk_node; 209 di_node_name_t di_node_name; 210 di_node_t root_node; 211 212 if (!DLLINK(h, di_init)) 213 break; 214 if (!DLLINK(h, di_fini)) 215 break; 216 if (!DLLINK(h, di_walk_node)) 217 break; 218 if (!DLLINK(h, di_node_name)) 219 break; 220 221 if ((root_node = (*di_init) ("/", DINFOSUBTREE)) != DI_NODE_NIL) { 222 (*di_walk_node) (root_node, DI_WALK_SIBFIRST, 223 di_node_name, walk_nodename); 224 (*di_fini) (root_node); 225 } 226 } while (0); 227 228 if (h) 229 dlclose(h); 230 } 231 232 #else 233 234 static sigjmp_buf common_jmp; 235 static void common_handler(int sig) 236 { 237 siglongjmp(common_jmp, sig); 238 } 239 240 #if defined(__sun) && defined(__SVR4) 241 # if defined(__GNUC__) && __GNUC__>=2 242 extern unsigned int getisax(unsigned int vec[], unsigned int sz) __attribute__ ((weak)); 243 # elif defined(__SUNPRO_C) 244 #pragma weak getisax 245 extern unsigned int getisax(unsigned int vec[], unsigned int sz); 246 # else 247 static unsigned int (*getisax) (unsigned int vec[], unsigned int sz) = NULL; 248 # endif 249 #endif 250 251 void OPENSSL_cpuid_setup(void) 252 { 253 char *e; 254 struct sigaction common_act, ill_oact, bus_oact; 255 sigset_t all_masked, oset; 256 static int trigger = 0; 257 258 if (trigger) 259 return; 260 trigger = 1; 261 262 if ((e = getenv("OPENSSL_sparcv9cap"))) { 263 OPENSSL_sparcv9cap_P[0] = strtoul(e, NULL, 0); 264 if ((e = strchr(e, ':'))) 265 OPENSSL_sparcv9cap_P[1] = strtoul(e + 1, NULL, 0); 266 return; 267 } 268 269 #if defined(__sun) && defined(__SVR4) 270 if (getisax != NULL) { 271 unsigned int vec[1]; 272 273 if (getisax (vec,1)) { 274 if (vec[0]&0x0020) OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS1; 275 if (vec[0]&0x0040) OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS2; 276 if (vec[0]&0x0080) OPENSSL_sparcv9cap_P[0] |= SPARCV9_BLK; 277 if (vec[0]&0x0100) OPENSSL_sparcv9cap_P[0] |= SPARCV9_FMADD; 278 if (vec[0]&0x0400) OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS3; 279 280 /* reconstruct %cfr copy */ 281 OPENSSL_sparcv9cap_P[1] = (vec[0]>>17)&0x3ff; 282 OPENSSL_sparcv9cap_P[1] |= (OPENSSL_sparcv9cap_P[1]&CFR_MONTMUL)<<1; 283 if (vec[0]&0x20000000) OPENSSL_sparcv9cap_P[1] |= CFR_CRC32C; 284 285 /* Some heuristics */ 286 /* all known VIS2-capable CPUs have unprivileged tick counter */ 287 if (OPENSSL_sparcv9cap_P[0]&SPARCV9_VIS2) 288 OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED; 289 290 OPENSSL_sparcv9cap_P[0] |= SPARCV9_PREFER_FPU; 291 292 /* detect UltraSPARC-Tx, see sparccpud.S for details... */ 293 if ((OPENSSL_sparcv9cap_P[0]&SPARCV9_VIS1) && 294 _sparcv9_vis1_instrument() >= 12) 295 OPENSSL_sparcv9cap_P[0] &= ~(SPARCV9_VIS1 | SPARCV9_PREFER_FPU); 296 } 297 298 if (sizeof(size_t) == 8) 299 OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK; 300 301 return; 302 } 303 #endif 304 305 /* Initial value, fits UltraSPARC-I&II... */ 306 OPENSSL_sparcv9cap_P[0] = SPARCV9_PREFER_FPU | SPARCV9_TICK_PRIVILEGED; 307 308 sigfillset(&all_masked); 309 sigdelset(&all_masked, SIGILL); 310 sigdelset(&all_masked, SIGTRAP); 311 # ifdef SIGEMT 312 sigdelset(&all_masked, SIGEMT); 313 # endif 314 sigdelset(&all_masked, SIGFPE); 315 sigdelset(&all_masked, SIGBUS); 316 sigdelset(&all_masked, SIGSEGV); 317 sigprocmask(SIG_SETMASK, &all_masked, &oset); 318 319 memset(&common_act, 0, sizeof(common_act)); 320 common_act.sa_handler = common_handler; 321 common_act.sa_mask = all_masked; 322 323 sigaction(SIGILL, &common_act, &ill_oact); 324 sigaction(SIGBUS, &common_act, &bus_oact); /* T1 fails 16-bit ldda [on 325 * Linux] */ 326 327 if (sigsetjmp(common_jmp, 1) == 0) { 328 _sparcv9_rdtick(); 329 OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED; 330 } 331 332 if (sigsetjmp(common_jmp, 1) == 0) { 333 _sparcv9_vis1_probe(); 334 OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS1 | SPARCV9_BLK; 335 /* detect UltraSPARC-Tx, see sparccpud.S for details... */ 336 if (_sparcv9_vis1_instrument() >= 12) 337 OPENSSL_sparcv9cap_P[0] &= ~(SPARCV9_VIS1 | SPARCV9_PREFER_FPU); 338 else { 339 _sparcv9_vis2_probe(); 340 OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS2; 341 } 342 } 343 344 if (sigsetjmp(common_jmp, 1) == 0) { 345 _sparcv9_fmadd_probe(); 346 OPENSSL_sparcv9cap_P[0] |= SPARCV9_FMADD; 347 } 348 349 /* 350 * VIS3 flag is tested independently from VIS1, unlike VIS2 that is, 351 * because VIS3 defines even integer instructions. 352 */ 353 if (sigsetjmp(common_jmp, 1) == 0) { 354 _sparcv9_vis3_probe(); 355 OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS3; 356 } 357 # if 0 /* was planned at some point but never 358 * implemented in hardware */ 359 if (sigsetjmp(common_jmp, 1) == 0) { 360 (void)_sparcv9_random(); 361 OPENSSL_sparcv9cap_P[0] |= SPARCV9_RANDOM; 362 } 363 # endif 364 365 /* 366 * In wait for better solution _sparcv9_rdcfr is masked by 367 * VIS3 flag, because it goes to uninterruptable endless 368 * loop on UltraSPARC II running Solaris. Things might be 369 * different on Linux... 370 */ 371 if ((OPENSSL_sparcv9cap_P[0] & SPARCV9_VIS3) && 372 sigsetjmp(common_jmp, 1) == 0) { 373 OPENSSL_sparcv9cap_P[1] = (unsigned int)_sparcv9_rdcfr(); 374 } 375 376 sigaction(SIGBUS, &bus_oact, NULL); 377 sigaction(SIGILL, &ill_oact, NULL); 378 379 sigprocmask(SIG_SETMASK, &oset, NULL); 380 381 if (sizeof(size_t) == 8) 382 OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK; 383 # ifdef __linux 384 else { 385 int ret = syscall(340); 386 387 if (ret >= 0 && ret & 1) 388 OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK; 389 } 390 # endif 391 } 392 393 #endif 394