mdw@git.distorted.org.uk Git - catacomb/blob - base/dispatch.c

   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 <assert.h>
  33 #include <ctype.h>
  34 #include <stdarg.h>
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <string.h>
  38
  39 #include <mLib/macros.h>
  40
  41 #include "dispatch.h"
  42
  43 /*----- Intel x86/AMD64 feature probing -----------------------------------*/
  44
  45 #if CPUFAM_X86 || CPUFAM_AMD64
  46
  47 enum {
  48   CPUID_1_D,                            /* eax = 1 => edx&?? */
  49 #  define CPUID1D_SSE2 (1u << 26)
  50 #  define CPUID1D_FXSR (1u << 24)
  51
  52   CPUID_1_C,                            /* eax = 1 => ecx&?? */
  53 #  define CPUID1C_PCLMUL (1u << 1)
  54 #  define CPUID1C_SSSE3 (1u << 9)
  55 #  define CPUID1C_AESNI (1u << 25)
  56 #  define CPUID1C_AVX (1u << 28)
  57 #  define CPUID1C_RDRAND (1u << 30)
  58
  59   CPUID_7_0_B,                          /* eax = 7, ecx = 0 => ebx&?? */
  60 #  define CPUID70B_RDSEED (1u << 18)
  61 };
  62
  63 struct cpuid { unsigned a, b, c, d; };
  64 extern int dispatch_x86ish_cpuid(struct cpuid *, unsigned a, unsigned c);
  65 extern int dispatch_x86ish_xmmregisters_p(void);
  66 extern int dispatch_x86ish_rdrand(unsigned op, unsigned *);
  67
  68 static void cpuid(struct cpuid *cc, unsigned a, unsigned c)
  69 {
  70   int rc = dispatch_x86ish_cpuid(cc, a, c);
  71   if (rc)
  72     dispatch_debug("CPUID instruction not available");
  73   else
  74     dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x",
  75                    a, c, cc->a, cc->b, cc->c, cc->d);
  76 }
  77
  78 static unsigned cpuid_maxleaf(void)
  79   { struct cpuid c; cpuid(&c, 0, 0); return (c.a); }
  80
  81 /* --- @cpuid_feature_p@ --- *
  82  *
  83  * Arguments:   @unsigned leaf@ = leaf to look up
  84  *              @unsigned bits@ = bits to check
  85  *
  86  * Returns:     Nonzero if all the requested bits are set in the requested
  87  *              CPUID result.
  88  */
  89
  90 static int cpuid_feature_p(unsigned leaf, unsigned bits)
  91 {
  92   struct cpuid c;
  93   unsigned r;
  94
  95   switch (leaf) {
  96     case CPUID_1_D:
  97       if (cpuid_maxleaf() < 1) return (0);
  98       cpuid(&c, 1, 0); r = c.d;
  99       break;
 100     case CPUID_1_C:
 101       if (cpuid_maxleaf() < 1) return (0);
 102       cpuid(&c, 1, 0); r = c.c;
 103       break;
 104     case CPUID_7_0_B:
 105       if (cpuid_maxleaf() < 7) return (0);
 106       cpuid(&c, 7, 0); r = c.b;
 107       break;
 108     default:
 109       assert(!"unknown cpuid leaf");
 110   }
 111   return ((r&bits) == bits);
 112 }
 113
 114 /* --- @xmm_registers_available_p@ --- *
 115  *
 116  * Arguments:   ---
 117  *
 118  * Returns:     Nonzero if the operating system has made the XMM registers
 119  *              available for use.
 120  */
 121
 122 static int xmm_registers_available_p(void)
 123 {
 124   int f = dispatch_x86ish_xmmregisters_p();
 125
 126   dispatch_debug("XMM registers %savailable", f ? "" : "not ");
 127   return (f);
 128 }
 129
 130 /* --- @rdrand_works_p@ --- *
 131  *
 132  *
 133  * Arguments:   ---
 134  *
 135  * Returns:     Nonzero if the `rdrand' instruction actually works.  Assumes
 136  *              that it's already been verified to be safe to issue.
 137  */
 138
 139 enum { OP_RDRAND, OP_RDSEED };
 140
 141 static int rdrand_works_p(unsigned op)
 142 {
 143   unsigned ref, x, i;
 144   const char *what;
 145
 146   switch (op) {
 147     case OP_RDRAND: what = "RDRAND"; break;
 148     case OP_RDSEED: what = "RDSEED"; break;
 149     default: assert(!"unexpected op");
 150   }
 151
 152   /* Check that it doesn't always give the same answer.  Try four times: this
 153    * will fail with probability %$2^{-128}$% with a truly random generator,
 154    * which seems fair enough.
 155    */
 156   if (dispatch_x86ish_rdrand(op, &ref)) goto fail;
 157   for (i = 0; i < 4; i++) {
 158     if (dispatch_x86ish_rdrand(op, &x)) goto fail;
 159     if (x != ref) goto not_stuck;
 160   }
 161   dispatch_debug("%s always returns 0x%08x!", what, ref);
 162   return (0);
 163
 164 not_stuck:
 165   dispatch_debug("%s instruction looks plausible", what);
 166   return (1);
 167
 168 fail:
 169   dispatch_debug("%s instruction fails too often", what);
 170   return (0);
 171 }
 172
 173 #endif
 174
 175 /*----- General feature probing using auxiliary vectors -------------------*/
 176
 177 /* Try to find the system's definitions for auxiliary vector entries. */
 178 #ifdef HAVE_SYS_AUXV_H
 179 #  include <sys/auxv.h>
 180 #endif
 181 #ifdef HAVE_LINUX_AUXVEC_H
 182 #  include <linux/auxvec.h>
 183 #endif
 184 #ifdef HAVE_ASM_HWCAP_H
 185 #  include <asm/hwcap.h>
 186 #endif
 187
 188 /* The type of entries in the auxiliary vector.  I'm assuming that `unsigned
 189  * long' matches each platform's word length; if this is false then we'll
 190  * need some host-specific tweaking here.
 191  */
 192 union auxval { long i; unsigned long u; const void *p; };
 193 struct auxentry { unsigned long type; union auxval value; };
 194
 195 /* Register each CPU family's interest in the auxiliary vector.  Make sure
 196  * that the necessary entry types are defined.  This is primarily ordered by
 197  * entry type to minimize duplication.
 198  */
 199 #if defined(AT_HWCAP) && CPUFAM_ARMEL
 200 #  define WANT_ANY 1
 201 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 202 #endif
 203
 204 #if defined(AT_HWCAP) && CPUFAM_ARM64
 205 #  define WANT_ANY 1
 206 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 207 #endif
 208
 209 #if defined(AT_HWCAP2) && CPUFAM_ARMEL
 210 #  define WANT_ANY 1
 211 #  define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
 212 #endif
 213
 214 /* If we couldn't find any interesting entries then we can switch all of this
 215  * machinery off.  Also do that if we have no means for atomic updates.
 216  */
 217 #if WANT_ANY && CPU_DISPATCH_P
 218
 219 /* The main output of this section is a bitmask of detected features.  The
 220  * least significant bit will be set if we've tried to probe.  Always access
 221  * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
 222  */
 223 static unsigned hwcaps = 0;
 224
 225 /* For each potentially interesting type which turned out not to exist or be
 226  * wanted, define a dummy macro for the sake of the next step.
 227  */
 228 #ifndef WANT_AT_HWCAP
 229 #  define WANT_AT_HWCAP(_)
 230 #endif
 231 #ifndef WANT_AT_HWCAP2
 232 #  define WANT_AT_HWCAP2(_)
 233 #endif
 234
 235 /* For each CPU family, define two lists.
 236  *
 237  *   * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
 238  *     family tried to register interest in above.  Each entry contains the
 239  *     interesting auxiliary vector entry type, the name of the union branch
 240  *     for its value, and the name of the slot in `struct auxprobe' in which
 241  *     to store the value.
 242  *
 243  *   * `CAPMAP' is a list describing the output features which the CPU family
 244  *     intends to satisfy from the auxiliary vector.  Each entry contains a
 245  *     feature name suffix, and the token name (for `check_env').
 246  */
 247 #if CPUFAM_ARMEL
 248 #  define WANTAUX(_)                                                    \
 249         WANT_AT_HWCAP(_)                                                \
 250         WANT_AT_HWCAP2(_)
 251 #  define CAPMAP(_)                                                     \
 252         _(ARM_VFP, "arm:vfp")                                           \
 253         _(ARM_NEON, "arm:neon")                                         \
 254         _(ARM_V4, "arm:v4")                                             \
 255         _(ARM_D32, "arm:d32")                                           \
 256         _(ARM_AES, "arm:aes")                                           \
 257         _(ARM_PMULL, "arm:pmull")
 258 #endif
 259 #if CPUFAM_ARM64
 260 #  define WANTAUX(_)                                                    \
 261         WANT_AT_HWCAP(_)
 262 #  define CAPMAP(_)                                                     \
 263         _(ARM_NEON, "arm:neon")                                         \
 264         _(ARM_AES, "arm:aes")                                           \
 265         _(ARM_PMULL, "arm:pmull")
 266 #endif
 267
 268 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
 269 enum {
 270   HFI_PROBED = 0,
 271 #define HFI__ENUM(feat, tok) HFI_##feat,
 272   CAPMAP(HFI__ENUM)
 273 #undef HFI__ENUM
 274   HFI__END
 275 };
 276 enum {
 277   HF_PROBED = 1,
 278 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
 279   CAPMAP(HF__FLAG)
 280 #undef HF__FLAG
 281   HF__END
 282 };
 283
 284 /* Build a structure in which we can capture the interesting data from the
 285  * auxiliary vector.
 286  */
 287 #define AUXUTYPE_i long
 288 #define AUXUTYPE_u unsigned long
 289 #define AUXUTYPE_p const void *
 290 struct auxprobe {
 291 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
 292   WANTAUX(AUXPROBE__SLOT)
 293 #undef AUXPROBE_SLOT
 294 };
 295
 296 /* --- @probe_hwcaps@ --- *
 297  *
 298  * Arguments:   ---
 299  *
 300  * Returns:     ---
 301  *
 302  * Use:         Attempt to find the auxiliary vector (which is well hidden)
 303  *              and discover interesting features from it.
 304  */
 305
 306 static void probe_hwcaps(void)
 307 {
 308   unsigned hw = HF_PROBED;
 309   struct auxprobe probed = { 0 };
 310
 311   /* Populate `probed' with the information we manage to retrieve from the
 312    * auxiliary vector.  Slots we couldn't find are left zero-valued.
 313    */
 314 #if defined(HAVE_GETAUXVAL)
 315   /* Shiny new libc lets us request individual entry types.  This is almost
 316    * too easy.
 317    */
 318 #  define CAP__GET(type, ubranch, slot)                                 \
 319         probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
 320   WANTAUX(CAP__GET)
 321 #else
 322   /* Otherwise we're a bit stuck, really.  Modern Linux kernels make a copy
 323    * of the vector available in `/procc' so we could try that.
 324    *
 325    * The usual place is stuck on the end of the environment vector, but that
 326    * may well have moved, and we have no way of telling whether it has or
 327    * whether there was ever an auxiliary vector there at all; so don't do
 328    * that.
 329    */
 330   {
 331     FILE *fp = 0;
 332     unsigned char *p = 0, *q = 0;
 333     const struct auxentry *a;
 334     size_t sz, off, n;
 335
 336     /* Open the file and read it into a memory chunk. */
 337     if ((fp = fopen("/proc/self/auxv", "rb")) == 0) goto clean;
 338     sz = 4096; off = 0;
 339     if ((p = malloc(sz)) == 0) goto clean;
 340     for (;;) {
 341       n = fread(p + off, 1, sz - off, fp);
 342       off += n;
 343       if (off < sz) break;
 344       sz *= 2; if ((q = realloc(p, sz)) == 0) break;
 345       p = q;
 346     }
 347
 348     /* Work through the vector (or as much of it as we found) and extract the
 349      * types we're interested in.
 350      */
 351     for (a = (const struct auxentry *)p,
 352            n = sz/sizeof(struct auxentry);
 353          n--; a++) {
 354       switch (a->type) {
 355 #define CAP__SWITCH(type, ubranch, slot)                                \
 356         case type: probed.slot = a->value.ubranch; break;
 357         WANTAUX(CAP__SWITCH)
 358         case AT_NULL: goto clean;
 359       }
 360     }
 361
 362   clean:
 363     if (p) free(p);
 364     if (fp) fclose(fp);
 365   }
 366 #endif
 367
 368   /* Each CPU family now has to pick through what was found and stashed in
 369    * `probed', and set the appropriate flag bits in `hw'.
 370    */
 371 #if CPUFAM_ARMEL
 372   if (probed.hwcap & HWCAP_VFPv3) hw |= HF_ARM_VFP;
 373   if (probed.hwcap & HWCAP_NEON) hw |= HF_ARM_NEON;
 374   if (probed.hwcap & HWCAP_VFPD32) hw |= HF_ARM_D32;
 375   if (probed.hwcap & HWCAP_VFPv4) hw |= HF_ARM_V4;
 376 #  ifdef HWCAP2_AES
 377   if (probed.hwcap2 & HWCAP2_AES) hw |= HF_ARM_AES;
 378 #  endif
 379 #  ifdef HWCAP2_PMULL
 380   if (probed.hwcap2 & HWCAP2_PMULL) hw |= HF_ARM_PMULL;
 381 #  endif
 382 #endif
 383 #if CPUFAM_ARM64
 384   if (probed.hwcap & HWCAP_ASIMD) hw |= HF_ARM_NEON;
 385   if (probed.hwcap & HWCAP_AES) hw |= HF_ARM_AES;
 386   if (probed.hwcap & HWCAP_PMULL) hw |= HF_ARM_PMULL;
 387 #endif
 388
 389   /* Store the bitmask of features we probed for everyone to see. */
 390   DISPATCH_STORE(hwcaps, hw);
 391
 392   /* Finally, make a report about the things we found.  (Doing this earlier
 393    * will pointlessly widen the window in which multiple threads will do the
 394    * above auxiliary-vector probing.)
 395    */
 396 #define CAP__DEBUG(feat, tok)                                           \
 397   dispatch_debug("check auxv for feature `%s': %s", tok,                \
 398                  hw & HF_##feat ? "available" : "absent");
 399   CAPMAP(CAP__DEBUG)
 400 #undef CAP__DEBUG
 401 }
 402
 403 /* --- @get_hwcaps@ --- *
 404  *
 405  * Arguments:   ---
 406  *
 407  * Returns:     A mask of hardware capabilities and other features, as probed
 408  *              from the auxiliary vector.
 409  */
 410
 411 static unsigned get_hwcaps(void)
 412 {
 413   unsigned hw;
 414
 415   DISPATCH_LOAD(hwcaps, hw);
 416   if (!(hwcaps & HF_PROBED)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps, hw); }
 417   return (hw);
 418 }
 419
 420 #endif
 421
 422 /*----- External interface ------------------------------------------------*/
 423
 424 /* --- @dispatch_debug@ --- *
 425  *
 426  * Arguments:   @const char *fmt@ = a format string
 427  *              @...@ = additional arguments
 428  *
 429  * Returns:     ---
 430  *
 431  * Use:         Writes a formatted message to standard output if dispatch
 432  *              debugging is enabled.
 433  */
 434
 435 void dispatch_debug(const char *fmt, ...)
 436 {
 437   va_list ap;
 438   const char *e = getenv("CATACOMB_CPUDISPATCH_DEBUG");
 439
 440   if (e && *e != 'n' && *e != '0') {
 441     va_start(ap, fmt);
 442     fputs("Catacomb CPUDISPATCH: ", stderr);
 443     vfprintf(stderr, fmt, ap);
 444     fputc('\n', stderr);
 445     va_end(ap);
 446   }
 447 }
 448
 449 /* --- @check_env@ --- *
 450  *
 451  * Arguments:   @const char *ftok@ = feature token
 452  *
 453  * Returns:     Zero if the feature is forced off; positive if it's forced
 454  *              on; negative if the user hasn't decided.
 455  *
 456  * Use:         Checks the environment variable `CATACOMB_CPUFEAT' for the
 457  *              feature token @ftok@.  The variable, if it exists, should be
 458  *              a space-separated sequence of `+tok' and `-tok' items.  These
 459  *              tokens may end in `*', which matches any suffix.
 460  */
 461
 462 static int IGNORABLE check_env(const char *ftok)
 463 {
 464   const char *p, *q, *pp;
 465   int d;
 466
 467   p = getenv("CATACOMB_CPUFEAT");
 468   if (!p) return (-1);
 469
 470   for (;;) {
 471     while (ISSPACE(*p)) p++;
 472     if (!*p) return (-1);
 473     switch (*p) {
 474       case '+': d = +1; p++; break;
 475       case '-': d =  0; p++; break;
 476       default:  d = -1;      break;
 477     }
 478     for (q = p; *q && !ISSPACE(*q); q++);
 479     if (d >= 0) {
 480       for (pp = ftok; p < q && *pp && *p == *pp; p++, pp++);
 481       if ((p == q && !*pp) || (*p == '*' && p + 1 == q)) return (d);
 482     }
 483     p = q;
 484   }
 485   return (-1);
 486 }
 487
 488 /* --- @cpu_feature_p@ --- *
 489  *
 490  * Arguments:   @unsigned feat@ = a @CPUFEAT_...@ code
 491  *
 492  * Returns:     Nonzero if the feature is available.
 493  */
 494
 495 #include <stdio.h>
 496
 497 static int IGNORABLE
 498   feat_debug(const char *ftok, const char *check, int verdict)
 499 {
 500   if (verdict >= 0) {
 501     dispatch_debug("feature `%s': %s -> %s", ftok, check,
 502                    verdict ? "available" : "absent");
 503   }
 504   return (verdict);
 505 }
 506
 507 int cpu_feature_p(int feat)
 508 {
 509   int IGNORABLE f;
 510   IGNORE(f);
 511 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat:             \
 512   if ((f = feat_debug(ftok, "environment override", check_env(ftok))) >= 0) \
 513     return (f);                                                         \
 514   else                                                                  \
 515     return (feat_debug(ftok, "runtime probe", cond));
 516
 517   switch (feat) {
 518 #if CPUFAM_X86 || CPUFAM_AMD64
 519     CASE_CPUFEAT(X86_SSE2, "x86:sse2",
 520                  cpuid_feature_p(CPUID_1_D, CPUID1D_SSE2) &&
 521                  xmm_registers_available_p());
 522     CASE_CPUFEAT(X86_AESNI, "x86:aesni",
 523                  cpuid_feature_p(CPUID_1_D, CPUID1C_AESNI) &&
 524                  xmm_registers_available_p());
 525     CASE_CPUFEAT(X86_RDRAND, "x86:rdrand",
 526                  cpuid_feature_p(CPUID_1_C, CPUID1C_RDRAND) &&
 527                  rdrand_works_p(OP_RDRAND));
 528     CASE_CPUFEAT(X86_AVX, "x86:avx",
 529                  cpuid_feature_p(CPUID_1_C, CPUID1C_AVX) &&
 530                  xmm_registers_available_p());
 531     CASE_CPUFEAT(X86_SSSE3, "x86:ssse3",
 532                  cpuid_feature_p(CPUID_1_C, CPUID1C_SSSE3) &&
 533                  xmm_registers_available_p());
 534     CASE_CPUFEAT(X86_PCLMUL, "x86:pclmul",
 535                  cpuid_feature_p(CPUID_1_C, CPUID1C_PCLMUL) &&
 536                  xmm_registers_available_p());
 537     CASE_CPUFEAT(X86_RDSEED, "x86:rdseed",
 538                  cpuid_feature_p(CPUID_7_0_B, CPUID70B_RDSEED) &&
 539                  rdrand_works_p(OP_RDSEED));
 540 #endif
 541 #ifdef CAPMAP
 542 #  define FEATP__CASE(feat, tok)                                        \
 543         CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
 544     CAPMAP(FEATP__CASE)
 545 #undef FEATP__CASE
 546 #endif
 547     default:
 548       dispatch_debug("denying unknown feature %d", feat);
 549       return (0);
 550   }
 551 #undef CASE_CPUFEAT
 552 }
 553
 554 /*----- That's all, folks -------------------------------------------------*/