| 1 | /* -*-c-*- |
| 2 | * |
| 3 | * CPU-specific dispatch |
| 4 | * |
| 5 | * (c) 2015 Straylight/Edgeware |
| 6 | */ |
| 7 | |
| 8 | /*----- Licensing notice --------------------------------------------------* |
| 9 | * |
| 10 | * This file is part of Catacomb. |
| 11 | * |
| 12 | * Catacomb is free software; you can redistribute it and/or modify |
| 13 | * it under the terms of the GNU Library General Public License as |
| 14 | * published by the Free Software Foundation; either version 2 of the |
| 15 | * License, or (at your option) any later version. |
| 16 | * |
| 17 | * Catacomb is distributed in the hope that it will be useful, |
| 18 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 19 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 20 | * GNU Library General Public License for more details. |
| 21 | * |
| 22 | * You should have received a copy of the GNU Library General Public |
| 23 | * License along with Catacomb; if not, write to the Free |
| 24 | * Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, |
| 25 | * MA 02111-1307, USA. |
| 26 | */ |
| 27 | |
| 28 | /*----- Header files ------------------------------------------------------*/ |
| 29 | |
| 30 | #include "config.h" |
| 31 | |
| 32 | #include <ctype.h> |
| 33 | #include <stdarg.h> |
| 34 | #include <stdio.h> |
| 35 | #include <stdlib.h> |
| 36 | #include <string.h> |
| 37 | |
| 38 | #include <mLib/macros.h> |
| 39 | |
| 40 | #include "dispatch.h" |
| 41 | |
| 42 | /*----- Intel x86/AMD64 feature probing -----------------------------------*/ |
| 43 | |
| 44 | #if CPUFAM_X86 || CPUFAM_AMD64 |
| 45 | |
| 46 | # define CPUID1D_SSE2 (1u << 26) |
| 47 | # define CPUID1D_FXSR (1u << 24) |
| 48 | # define CPUID1C_PCLMUL (1u << 1) |
| 49 | # define CPUID1C_SSSE3 (1u << 9) |
| 50 | # define CPUID1C_AESNI (1u << 25) |
| 51 | # define CPUID1C_AVX (1u << 28) |
| 52 | # define CPUID1C_RDRAND (1u << 30) |
| 53 | |
| 54 | struct cpuid { unsigned a, b, c, d; }; |
| 55 | extern int dispatch_x86ish_cpuid(struct cpuid *, unsigned a, unsigned c); |
| 56 | extern int dispatch_x86ish_xmmregisters_p(void); |
| 57 | extern int dispatch_x86ish_rdrand(unsigned *); |
| 58 | |
| 59 | static void cpuid(struct cpuid *cc, unsigned a, unsigned c) |
| 60 | { |
| 61 | int rc = dispatch_x86ish_cpuid(cc, a, c); |
| 62 | if (rc) |
| 63 | dispatch_debug("CPUID instruction not available"); |
| 64 | else |
| 65 | dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x", |
| 66 | a, c, cc->a, cc->b, cc->c, cc->d); |
| 67 | } |
| 68 | |
| 69 | static unsigned cpuid_maxleaf(void) |
| 70 | { struct cpuid c; cpuid(&c, 0, 0); return (c.a); } |
| 71 | |
| 72 | /* --- @cpuid_features_p@ --- * |
| 73 | * |
| 74 | * Arguments: @unsigned dbits@ = bits to check in EDX |
| 75 | * @unsigned cbits@ = bits to check in ECX |
| 76 | * |
| 77 | * Returns: Nonzero if all the requested bits are set in the CPUID result |
| 78 | * on leaf 1. |
| 79 | */ |
| 80 | |
| 81 | static int cpuid_features_p(unsigned dbits, unsigned cbits) |
| 82 | { |
| 83 | struct cpuid c; |
| 84 | if (cpuid_maxleaf() < 1) return (0); |
| 85 | cpuid(&c, 1, 0); |
| 86 | return ((c.d & dbits) == dbits && (c.c & cbits) == cbits); |
| 87 | } |
| 88 | |
| 89 | /* --- @xmm_registers_available_p@ --- * |
| 90 | * |
| 91 | * Arguments: --- |
| 92 | * |
| 93 | * Returns: Nonzero if the operating system has made the XMM registers |
| 94 | * available for use. |
| 95 | */ |
| 96 | |
| 97 | static int xmm_registers_available_p(void) |
| 98 | { |
| 99 | int f = dispatch_x86ish_xmmregisters_p(); |
| 100 | |
| 101 | dispatch_debug("XMM registers %savailable", f ? "" : "not "); |
| 102 | return (f); |
| 103 | } |
| 104 | |
| 105 | /* --- @rdrand_works_p@ --- * |
| 106 | * |
| 107 | * |
| 108 | * Arguments: --- |
| 109 | * |
| 110 | * Returns: Nonzero if the `rdrand' instruction actually works. Assumes |
| 111 | * that it's already been verified to be safe to issue. |
| 112 | */ |
| 113 | |
| 114 | static int rdrand_works_p(void) |
| 115 | { |
| 116 | unsigned ref, x, i; |
| 117 | |
| 118 | /* Check that it doesn't always give the same answer. Try four times: this |
| 119 | * will fail with probability %$2^{-128}$% with a truly random generator, |
| 120 | * which seems fair enough. |
| 121 | */ |
| 122 | if (dispatch_x86ish_rdrand(&ref)) goto fail; |
| 123 | for (i = 0; i < 4; i++) { |
| 124 | if (dispatch_x86ish_rdrand(&x)) goto fail; |
| 125 | if (x != ref) goto not_stuck; |
| 126 | } |
| 127 | dispatch_debug("RDRAND always returns 0x%08x!", ref); |
| 128 | return (0); |
| 129 | |
| 130 | not_stuck: |
| 131 | dispatch_debug("RDRAND instruction looks plausible"); |
| 132 | return (1); |
| 133 | |
| 134 | fail: |
| 135 | dispatch_debug("RDRAND instruction fails too often"); |
| 136 | return (0); |
| 137 | } |
| 138 | |
| 139 | #endif |
| 140 | |
| 141 | /*----- General feature probing using auxiliary vectors -------------------*/ |
| 142 | |
| 143 | /* Try to find the system's definitions for auxiliary vector entries. */ |
| 144 | #ifdef HAVE_SYS_AUXV_H |
| 145 | # include <sys/auxv.h> |
| 146 | #endif |
| 147 | #ifdef HAVE_LINUX_AUXVEC_H |
| 148 | # include <linux/auxvec.h> |
| 149 | #endif |
| 150 | #ifdef HAVE_ASM_HWCAP_H |
| 151 | # include <asm/hwcap.h> |
| 152 | #endif |
| 153 | |
| 154 | /* The type of entries in the auxiliary vector. I'm assuming that `unsigned |
| 155 | * long' matches each platform's word length; if this is false then we'll |
| 156 | * need some host-specific tweaking here. |
| 157 | */ |
| 158 | union auxval { long i; unsigned long u; const void *p; }; |
| 159 | struct auxentry { unsigned long type; union auxval value; }; |
| 160 | |
| 161 | /* Register each CPU family's interest in the auxiliary vector. Make sure |
| 162 | * that the necessary entry types are defined. This is primarily ordered by |
| 163 | * entry type to minimize duplication. |
| 164 | */ |
| 165 | #if defined(AT_HWCAP) && CPUFAM_ARMEL |
| 166 | # define WANT_ANY 1 |
| 167 | # define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap) |
| 168 | #endif |
| 169 | |
| 170 | #if defined(AT_HWCAP) && CPUFAM_ARM64 |
| 171 | # define WANT_ANY 1 |
| 172 | # define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap) |
| 173 | #endif |
| 174 | |
| 175 | #if defined(AT_HWCAP2) && CPUFAM_ARMEL |
| 176 | # define WANT_ANY 1 |
| 177 | # define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2) |
| 178 | #endif |
| 179 | |
| 180 | /* If we couldn't find any interesting entries then we can switch all of this |
| 181 | * machinery off. Also do that if we have no means for atomic updates. |
| 182 | */ |
| 183 | #if WANT_ANY && CPU_DISPATCH_P |
| 184 | |
| 185 | /* The main output of this section is a bitmask of detected features. The |
| 186 | * least significant bit will be set if we've tried to probe. Always access |
| 187 | * this using `DISPATCH_LOAD' and `DISPATCH_STORE'. |
| 188 | */ |
| 189 | static unsigned hwcaps = 0; |
| 190 | |
| 191 | /* For each potentially interesting type which turned out not to exist or be |
| 192 | * wanted, define a dummy macro for the sake of the next step. |
| 193 | */ |
| 194 | #ifndef WANT_AT_HWCAP |
| 195 | # define WANT_AT_HWCAP(_) |
| 196 | #endif |
| 197 | #ifndef WANT_AT_HWCAP2 |
| 198 | # define WANT_AT_HWCAP2(_) |
| 199 | #endif |
| 200 | |
| 201 | /* For each CPU family, define two lists. |
| 202 | * |
| 203 | * * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU |
| 204 | * family tried to register interest in above. Each entry contains the |
| 205 | * interesting auxiliary vector entry type, the name of the union branch |
| 206 | * for its value, and the name of the slot in `struct auxprobe' in which |
| 207 | * to store the value. |
| 208 | * |
| 209 | * * `CAPMAP' is a list describing the output features which the CPU family |
| 210 | * intends to satisfy from the auxiliary vector. Each entry contains a |
| 211 | * feature name suffix, and the token name (for `check_env'). |
| 212 | */ |
| 213 | #if CPUFAM_ARMEL |
| 214 | # define WANTAUX(_) \ |
| 215 | WANT_AT_HWCAP(_) \ |
| 216 | WANT_AT_HWCAP2(_) |
| 217 | # define CAPMAP(_) \ |
| 218 | _(ARM_VFP, "arm:vfp") \ |
| 219 | _(ARM_NEON, "arm:neon") \ |
| 220 | _(ARM_V4, "arm:v4") \ |
| 221 | _(ARM_D32, "arm:d32") \ |
| 222 | _(ARM_AES, "arm:aes") \ |
| 223 | _(ARM_PMULL, "arm:pmull") |
| 224 | #endif |
| 225 | #if CPUFAM_ARM64 |
| 226 | # define WANTAUX(_) \ |
| 227 | WANT_AT_HWCAP(_) |
| 228 | # define CAPMAP(_) \ |
| 229 | _(ARM_NEON, "arm:neon") \ |
| 230 | _(ARM_AES, "arm:aes") \ |
| 231 | _(ARM_PMULL, "arm:pmull") |
| 232 | #endif |
| 233 | |
| 234 | /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */ |
| 235 | enum { |
| 236 | HFI_PROBED = 0, |
| 237 | #define HFI__ENUM(feat, tok) HFI_##feat, |
| 238 | CAPMAP(HFI__ENUM) |
| 239 | #undef HFI__ENUM |
| 240 | HFI__END |
| 241 | }; |
| 242 | enum { |
| 243 | HF_PROBED = 1, |
| 244 | #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat, |
| 245 | CAPMAP(HF__FLAG) |
| 246 | #undef HF__FLAG |
| 247 | HF__END |
| 248 | }; |
| 249 | |
| 250 | /* Build a structure in which we can capture the interesting data from the |
| 251 | * auxiliary vector. |
| 252 | */ |
| 253 | #define AUXUTYPE_i long |
| 254 | #define AUXUTYPE_u unsigned long |
| 255 | #define AUXUTYPE_p const void * |
| 256 | struct auxprobe { |
| 257 | #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot; |
| 258 | WANTAUX(AUXPROBE__SLOT) |
| 259 | #undef AUXPROBE_SLOT |
| 260 | }; |
| 261 | |
| 262 | /* --- @probe_hwcaps@ --- * |
| 263 | * |
| 264 | * Arguments: --- |
| 265 | * |
| 266 | * Returns: --- |
| 267 | * |
| 268 | * Use: Attempt to find the auxiliary vector (which is well hidden) |
| 269 | * and discover interesting features from it. |
| 270 | */ |
| 271 | |
| 272 | static void probe_hwcaps(void) |
| 273 | { |
| 274 | unsigned hw = HF_PROBED; |
| 275 | struct auxprobe probed = { 0 }; |
| 276 | |
| 277 | /* Populate `probed' with the information we manage to retrieve from the |
| 278 | * auxiliary vector. Slots we couldn't find are left zero-valued. |
| 279 | */ |
| 280 | #if defined(HAVE_GETAUXVAL) |
| 281 | /* Shiny new libc lets us request individual entry types. This is almost |
| 282 | * too easy. |
| 283 | */ |
| 284 | # define CAP__GET(type, ubranch, slot) \ |
| 285 | probed.slot = (AUXUTYPE_##ubranch)getauxval(type); |
| 286 | WANTAUX(CAP__GET) |
| 287 | #else |
| 288 | /* Otherwise we're a bit stuck, really. Modern Linux kernels make a copy |
| 289 | * of the vector available in `/procc' so we could try that. |
| 290 | * |
| 291 | * The usual place is stuck on the end of the environment vector, but that |
| 292 | * may well have moved, and we have no way of telling whether it has or |
| 293 | * whether there was ever an auxiliary vector there at all; so don't do |
| 294 | * that. |
| 295 | */ |
| 296 | { |
| 297 | FILE *fp = 0; |
| 298 | unsigned char *p = 0, *q = 0; |
| 299 | const struct auxentry *a; |
| 300 | size_t sz, off, n; |
| 301 | |
| 302 | /* Open the file and read it into a memory chunk. */ |
| 303 | if ((fp = fopen("/proc/self/auxv", "rb")) == 0) goto clean; |
| 304 | sz = 4096; off = 0; |
| 305 | if ((p = malloc(sz)) == 0) goto clean; |
| 306 | for (;;) { |
| 307 | n = fread(p + off, 1, sz - off, fp); |
| 308 | off += n; |
| 309 | if (off < sz) break; |
| 310 | sz *= 2; if ((q = realloc(p, sz)) == 0) break; |
| 311 | p = q; |
| 312 | } |
| 313 | |
| 314 | /* Work through the vector (or as much of it as we found) and extract the |
| 315 | * types we're interested in. |
| 316 | */ |
| 317 | for (a = (const struct auxentry *)p, |
| 318 | n = sz/sizeof(struct auxentry); |
| 319 | n--; a++) { |
| 320 | switch (a->type) { |
| 321 | #define CAP__SWITCH(type, ubranch, slot) \ |
| 322 | case type: probed.slot = a->value.ubranch; break; |
| 323 | WANTAUX(CAP__SWITCH) |
| 324 | case AT_NULL: goto clean; |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | clean: |
| 329 | if (p) free(p); |
| 330 | if (fp) fclose(fp); |
| 331 | } |
| 332 | #endif |
| 333 | |
| 334 | /* Each CPU family now has to pick through what was found and stashed in |
| 335 | * `probed', and set the appropriate flag bits in `hw'. |
| 336 | */ |
| 337 | #if CPUFAM_ARMEL |
| 338 | if (probed.hwcap & HWCAP_VFPv3) hw |= HF_ARM_VFP; |
| 339 | if (probed.hwcap & HWCAP_NEON) hw |= HF_ARM_NEON; |
| 340 | if (probed.hwcap & HWCAP_VFPD32) hw |= HF_ARM_D32; |
| 341 | if (probed.hwcap & HWCAP_VFPv4) hw |= HF_ARM_V4; |
| 342 | # ifdef HWCAP2_AES |
| 343 | if (probed.hwcap2 & HWCAP2_AES) hw |= HF_ARM_AES; |
| 344 | # endif |
| 345 | # ifdef HWCAP2_PMULL |
| 346 | if (probed.hwcap2 & HWCAP2_PMULL) hw |= HF_ARM_PMULL; |
| 347 | # endif |
| 348 | #endif |
| 349 | #if CPUFAM_ARM64 |
| 350 | if (probed.hwcap & HWCAP_ASIMD) hw |= HF_ARM_NEON; |
| 351 | if (probed.hwcap & HWCAP_AES) hw |= HF_ARM_AES; |
| 352 | if (probed.hwcap & HWCAP_PMULL) hw |= HF_ARM_PMULL; |
| 353 | #endif |
| 354 | |
| 355 | /* Store the bitmask of features we probed for everyone to see. */ |
| 356 | DISPATCH_STORE(hwcaps, hw); |
| 357 | |
| 358 | /* Finally, make a report about the things we found. (Doing this earlier |
| 359 | * will pointlessly widen the window in which multiple threads will do the |
| 360 | * above auxiliary-vector probing.) |
| 361 | */ |
| 362 | #define CAP__DEBUG(feat, tok) \ |
| 363 | dispatch_debug("check auxv for feature `%s': %s", tok, \ |
| 364 | hw & HF_##feat ? "available" : "absent"); |
| 365 | CAPMAP(CAP__DEBUG) |
| 366 | #undef CAP__DEBUG |
| 367 | } |
| 368 | |
| 369 | /* --- @get_hwcaps@ --- * |
| 370 | * |
| 371 | * Arguments: --- |
| 372 | * |
| 373 | * Returns: A mask of hardware capabilities and other features, as probed |
| 374 | * from the auxiliary vector. |
| 375 | */ |
| 376 | |
| 377 | static unsigned get_hwcaps(void) |
| 378 | { |
| 379 | unsigned hw; |
| 380 | |
| 381 | DISPATCH_LOAD(hwcaps, hw); |
| 382 | if (!(hwcaps & HF_PROBED)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps, hw); } |
| 383 | return (hw); |
| 384 | } |
| 385 | |
| 386 | #endif |
| 387 | |
| 388 | /*----- External interface ------------------------------------------------*/ |
| 389 | |
| 390 | /* --- @dispatch_debug@ --- * |
| 391 | * |
| 392 | * Arguments: @const char *fmt@ = a format string |
| 393 | * @...@ = additional arguments |
| 394 | * |
| 395 | * Returns: --- |
| 396 | * |
| 397 | * Use: Writes a formatted message to standard output if dispatch |
| 398 | * debugging is enabled. |
| 399 | */ |
| 400 | |
| 401 | void dispatch_debug(const char *fmt, ...) |
| 402 | { |
| 403 | va_list ap; |
| 404 | const char *e = getenv("CATACOMB_CPUDISPATCH_DEBUG"); |
| 405 | |
| 406 | if (e && *e != 'n' && *e != '0') { |
| 407 | va_start(ap, fmt); |
| 408 | fputs("Catacomb CPUDISPATCH: ", stderr); |
| 409 | vfprintf(stderr, fmt, ap); |
| 410 | fputc('\n', stderr); |
| 411 | va_end(ap); |
| 412 | } |
| 413 | } |
| 414 | |
| 415 | /* --- @check_env@ --- * |
| 416 | * |
| 417 | * Arguments: @const char *ftok@ = feature token |
| 418 | * |
| 419 | * Returns: Zero if the feature is forced off; positive if it's forced |
| 420 | * on; negative if the user hasn't decided. |
| 421 | * |
| 422 | * Use: Checks the environment variable `CATACOMB_CPUFEAT' for the |
| 423 | * feature token @ftok@. The variable, if it exists, should be |
| 424 | * a space-separated sequence of `+tok' and `-tok' items. These |
| 425 | * tokens may end in `*', which matches any suffix. |
| 426 | */ |
| 427 | |
| 428 | static int IGNORABLE check_env(const char *ftok) |
| 429 | { |
| 430 | const char *p, *q, *pp; |
| 431 | int d; |
| 432 | |
| 433 | p = getenv("CATACOMB_CPUFEAT"); |
| 434 | if (!p) return (-1); |
| 435 | |
| 436 | for (;;) { |
| 437 | while (ISSPACE(*p)) p++; |
| 438 | if (!*p) return (-1); |
| 439 | switch (*p) { |
| 440 | case '+': d = +1; p++; break; |
| 441 | case '-': d = 0; p++; break; |
| 442 | default: d = -1; break; |
| 443 | } |
| 444 | for (q = p; *q && !ISSPACE(*q); q++); |
| 445 | if (d >= 0) { |
| 446 | for (pp = ftok; p < q && *pp && *p == *pp; p++, pp++); |
| 447 | if ((p == q && !*pp) || (*p == '*' && p + 1 == q)) return (d); |
| 448 | } |
| 449 | p = q; |
| 450 | } |
| 451 | return (-1); |
| 452 | } |
| 453 | |
| 454 | /* --- @cpu_feature_p@ --- * |
| 455 | * |
| 456 | * Arguments: @unsigned feat@ = a @CPUFEAT_...@ code |
| 457 | * |
| 458 | * Returns: Nonzero if the feature is available. |
| 459 | */ |
| 460 | |
| 461 | #include <stdio.h> |
| 462 | |
| 463 | static int IGNORABLE |
| 464 | feat_debug(const char *ftok, const char *check, int verdict) |
| 465 | { |
| 466 | if (verdict >= 0) { |
| 467 | dispatch_debug("feature `%s': %s -> %s", ftok, check, |
| 468 | verdict ? "available" : "absent"); |
| 469 | } |
| 470 | return (verdict); |
| 471 | } |
| 472 | |
| 473 | int cpu_feature_p(int feat) |
| 474 | { |
| 475 | int IGNORABLE f; |
| 476 | IGNORE(f); |
| 477 | #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat: \ |
| 478 | if ((f = feat_debug(ftok, "environment override", check_env(ftok))) >= 0) \ |
| 479 | return (f); \ |
| 480 | else \ |
| 481 | return (feat_debug(ftok, "runtime probe", cond)); |
| 482 | |
| 483 | switch (feat) { |
| 484 | #if CPUFAM_X86 || CPUFAM_AMD64 |
| 485 | CASE_CPUFEAT(X86_SSE2, "x86:sse2", |
| 486 | cpuid_features_p(CPUID1D_SSE2, 0) && |
| 487 | xmm_registers_available_p()); |
| 488 | CASE_CPUFEAT(X86_AESNI, "x86:aesni", |
| 489 | cpuid_features_p(CPUID1D_SSE2, CPUID1C_AESNI) && |
| 490 | xmm_registers_available_p()); |
| 491 | CASE_CPUFEAT(X86_RDRAND, "x86:rdrand", |
| 492 | cpuid_features_p(0, CPUID1C_RDRAND) && rdrand_works_p()); |
| 493 | CASE_CPUFEAT(X86_AVX, "x86:avx", |
| 494 | cpuid_features_p(0, CPUID1C_AVX) && |
| 495 | xmm_registers_available_p()); |
| 496 | CASE_CPUFEAT(X86_SSSE3, "x86:ssse3", |
| 497 | cpuid_features_p(0, CPUID1C_SSSE3) && |
| 498 | xmm_registers_available_p()); |
| 499 | CASE_CPUFEAT(X86_PCLMUL, "x86:pclmul", |
| 500 | cpuid_features_p(0, CPUID1C_PCLMUL) && |
| 501 | xmm_registers_available_p()); |
| 502 | #endif |
| 503 | #ifdef CAPMAP |
| 504 | # define FEATP__CASE(feat, tok) \ |
| 505 | CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat) |
| 506 | CAPMAP(FEATP__CASE) |
| 507 | #undef FEATP__CASE |
| 508 | #endif |
| 509 | default: |
| 510 | dispatch_debug("denying unknown feature %d", feat); |
| 511 | return (0); |
| 512 | } |
| 513 | #undef CASE_CPUFEAT |
| 514 | } |
| 515 | |
| 516 | /*----- That's all, folks -------------------------------------------------*/ |