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 <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 EFLAGS_ID (1u << 21)
  47 #  define CPUID1D_SSE2 (1u << 26)
  48 #  define CPUID1D_FXSR (1u << 24)
  49 #  define CPUID1C_PCLMUL (1u << 1)
  50 #  define CPUID1C_SSSE3 (1u << 9)
  51 #  define CPUID1C_AESNI (1u << 25)
  52 #  define CPUID1C_AVX (1u << 28)
  53 #  define CPUID1C_RDRAND (1u << 30)
  54
  55 struct cpuid { unsigned a, b, c, d; };
  56
  57 /* --- @cpuid@ --- *
  58  *
  59  * Arguments:   @struct cpuid *cc@ = where to write the result
  60  *              @unsigned a, c@ = EAX and ECX registers to set
  61  *
  62  * Returns:     ---
  63  *
  64  * Use:         Minimal C wrapper around the x86 `CPUID' instruction.  Checks
  65  *              that the instruction is actually available before invoking
  66  *              it; fills the output structure with zero if it's not going to
  67  *              work.
  68  */
  69
  70 #ifdef __GNUC__
  71 #  if CPUFAM_X86
  72 static __inline__ unsigned getflags(void)
  73   { unsigned f; __asm__ ("pushf; popl %0" : "=g" (f)); return (f); }
  74 static __inline__ unsigned setflags(unsigned f)
  75 {
  76   unsigned ff;
  77   __asm__ ("pushf; pushl %1; popf; pushf; popl %0; popf"
  78            : "=r" (ff)
  79            : "r" (f));
  80   return (ff);
  81 }
  82 #  else
  83 static __inline__ unsigned long getflags(void)
  84   { unsigned long f; __asm__ ("pushf; popq %0" : "=g" (f)); return (f); }
  85 static __inline__ unsigned long long setflags(unsigned long f)
  86 {
  87   unsigned long ff;
  88   __asm__ ("pushf; pushq %1; popf; pushf; popq %0; popf"
  89            : "=r" (ff)
  90            : "r" (f));
  91   return (ff);
  92 }
  93 #  endif
  94 #endif
  95
  96 static void cpuid(struct cpuid *cc, unsigned a, unsigned c)
  97 {
  98 #ifdef __GNUC__
  99   unsigned f;
 100 #endif
 101
 102   cc->a = cc->b = cc->c = cc->d = 0;
 103
 104 #ifdef __GNUC__
 105   /* Stupid dance to detect whether the CPUID instruction is available. */
 106   f = getflags();
 107   if (!(setflags(f |  EFLAGS_ID) & EFLAGS_ID) ||
 108         setflags(f & ~EFLAGS_ID) & EFLAGS_ID) {
 109     dispatch_debug("CPUID instruction not available");
 110     return;
 111   }
 112   setflags(f);
 113
 114   /* Alas, EBX is magical in PIC code, so abuse ESI instead.  This isn't
 115    * pretty, but it works.
 116    */
 117 #  if CPUFAM_X86
 118   __asm__ ("pushl %%ebx; cpuid; movl %%ebx, %%esi; popl %%ebx"
 119            : "=a" (cc->a), "=S" (cc->b), "=c" (cc->c), "=d" (cc->d)
 120            : "a" (a) , "c" (c));
 121 #  elif CPUFAM_AMD64
 122   __asm__ ("pushq %%rbx; cpuid; movl %%ebx, %%esi; popq %%rbx"
 123            : "=a" (cc->a), "=S" (cc->b), "=c" (cc->c), "=d" (cc->d)
 124            : "a" (a) , "c" (c));
 125 #  else
 126 #    error "I'm confused."
 127 #  endif
 128   dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x",
 129                  a, c, cc->a, cc->b, cc->c, cc->d);
 130 #else
 131   dispatch_debug("GNU inline assembler not available; can't CPUID");
 132 #endif
 133 }
 134
 135 static unsigned cpuid_maxleaf(void)
 136   { struct cpuid c; cpuid(&c, 0, 0); return (c.a); }
 137
 138 /* --- @cpuid_features_p@ --- *
 139  *
 140  * Arguments:   @unsigned dbits@ = bits to check in EDX
 141  *              @unsigned cbits@ = bits to check in ECX
 142  *
 143  * Returns:     Nonzero if all the requested bits are set in the CPUID result
 144  *              on leaf 1.
 145  */
 146
 147 static int cpuid_features_p(unsigned dbits, unsigned cbits)
 148 {
 149   struct cpuid c;
 150   if (cpuid_maxleaf() < 1) return (0);
 151   cpuid(&c, 1, 0);
 152   return ((c.d & dbits) == dbits && (c.c & cbits) == cbits);
 153 }
 154
 155 /* --- @xmm_registers_available_p@ --- *
 156  *
 157  * Arguments:   ---
 158  *
 159  * Returns:     Nonzero if the operating system has made the XMM registers
 160  *              available for use.
 161  */
 162
 163 static int xmm_registers_available_p(void)
 164 {
 165 #ifdef __GNUC__
 166   unsigned f;
 167   /* This hack is by Agner Fog.  Use FXSAVE/FXRSTOR to figure out whether the
 168    * XMM registers are actually alive.
 169    */
 170   if (!cpuid_features_p(CPUID1D_FXSR, 0)) return (0);
 171 #  if CPUFAM_X86
 172   __asm__ ("movl %%esp, %%edx; subl $512, %%esp; andl $~15, %%esp\n"
 173            "fxsave (%%esp)\n"
 174            "movl 160(%%esp), %%eax; xorl $0xaaaa5555, 160(%%esp)\n"
 175            "fxrstor (%%esp); fxsave (%%esp)\n"
 176            "movl 160(%%esp), %%ecx; movl %%eax, 160(%%esp)\n"
 177            "fxrstor (%%esp); movl %%edx, %%esp\n"
 178            "xorl %%ecx, %%eax"
 179            : "=a" (f)
 180            : /* no inputs */
 181            : "%ecx", "%edx");
 182 #  elif CPUFAM_AMD64
 183   __asm__ ("movq %%rsp, %%rdx; subq $512, %%rsp; andq $~15, %%rsp\n"
 184            "fxsave (%%rsp)\n"
 185            "movl 160(%%rsp), %%eax; xorl $0xaaaa5555, 160(%%rsp)\n"
 186            "fxrstor (%%rsp); fxsave (%%rsp)\n"
 187            "movl 160(%%rsp), %%ecx; movl %%eax, 160(%%rsp)\n"
 188            "fxrstor (%%rsp); movq %%rdx, %%rsp\n"
 189            "xorl %%ecx, %%eax"
 190            : "=a" (f)
 191            : /* no inputs */
 192            : "%ecx", "%rdx");
 193 #  else
 194 #    error "I'm confused."
 195 #  endif
 196   dispatch_debug("XMM registers %savailable", f ? "" : "not ");
 197   return (f);
 198 #else
 199   dispatch_debug("GNU inline assembler not available; can't check for XMM");
 200   return (0);
 201 #endif
 202 }
 203
 204 #endif
 205
 206 /*----- General feature probing using auxiliary vectors -------------------*/
 207
 208 /* Try to find the system's definitions for auxiliary vector entries. */
 209 #ifdef HAVE_SYS_AUXV_H
 210 #  include <sys/auxv.h>
 211 #endif
 212 #ifdef HAVE_LINUX_AUXVEC_H
 213 #  include <linux/auxvec.h>
 214 #endif
 215 #ifdef HAVE_ASM_HWCAP_H
 216 #  include <asm/hwcap.h>
 217 #endif
 218
 219 /* The type of entries in the auxiliary vector.  I'm assuming that `unsigned
 220  * long' matches each platform's word length; if this is false then we'll
 221  * need some host-specific tweaking here.
 222  */
 223 union auxval { long i; unsigned long u; const void *p; };
 224 struct auxentry { unsigned long type; union auxval value; };
 225
 226 /* Register each CPU family's interest in the auxiliary vector.  Make sure
 227  * that the necessary entry types are defined.  This is primarily ordered by
 228  * entry type to minimize duplication.
 229  */
 230 #if defined(AT_HWCAP) && CPUFAM_ARMEL
 231 #  define WANT_ANY 1
 232 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 233 #endif
 234
 235 #if defined(AT_HWCAP) && CPUFAM_ARM64
 236 #  define WANT_ANY 1
 237 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 238 #endif
 239
 240 #if defined(AT_HWCAP2) && CPUFAM_ARMEL
 241 #  define WANT_ANY 1
 242 #  define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
 243 #endif
 244
 245 /* If we couldn't find any interesting entries then we can switch all of this
 246  * machinery off.  Also do that if we have no means for atomic updates.
 247  */
 248 #if WANT_ANY && CPU_DISPATCH_P
 249
 250 /* The main output of this section is a bitmask of detected features.  The
 251  * least significant bit will be set if we've tried to probe.  Always access
 252  * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
 253  */
 254 static unsigned hwcaps = 0;
 255
 256 /* For each potentially interesting type which turned out not to exist or be
 257  * wanted, define a dummy macro for the sake of the next step.
 258  */
 259 #ifndef WANT_AT_HWCAP
 260 #  define WANT_AT_HWCAP(_)
 261 #endif
 262 #ifndef WANT_AT_HWCAP2
 263 #  define WANT_AT_HWCAP2(_)
 264 #endif
 265
 266 /* For each CPU family, define two lists.
 267  *
 268  *   * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
 269  *     family tried to register interest in above.  Each entry contains the
 270  *     interesting auxiliary vector entry type, the name of the union branch
 271  *     for its value, and the name of the slot in `struct auxprobe' in which
 272  *     to store the value.
 273  *
 274  *   * `CAPMAP' is a list describing the output features which the CPU family
 275  *     intends to satisfy from the auxiliary vector.  Each entry contains a
 276  *     feature name suffix, and the token name (for `check_env').
 277  */
 278 #if CPUFAM_ARMEL
 279 #  define WANTAUX(_)                                                    \
 280         WANT_AT_HWCAP(_)                                                \
 281         WANT_AT_HWCAP2(_)
 282 #  define CAPMAP(_)                                                     \
 283         _(ARM_VFP, "arm:vfp")                                           \
 284         _(ARM_NEON, "arm:neon")                                         \
 285         _(ARM_V4, "arm:v4")                                             \
 286         _(ARM_D32, "arm:d32")                                           \
 287         _(ARM_AES, "arm:aes")                                           \
 288         _(ARM_PMULL, "arm:pmull")
 289 #endif
 290 #if CPUFAM_ARM64
 291 #  define WANTAUX(_)                                                    \
 292         WANT_AT_HWCAP(_)
 293 #  define CAPMAP(_)                                                     \
 294         _(ARM_AES, "arm:aes")                                           \
 295         _(ARM_PMULL, "arm:pmull")
 296 #endif
 297
 298 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
 299 enum {
 300   HFI_PROBED = 0,
 301 #define HFI__ENUM(feat, tok) HFI_##feat,
 302   CAPMAP(HFI__ENUM)
 303 #undef HFI__ENUM
 304   HFI__END
 305 };
 306 enum {
 307   HF_PROBED = 1,
 308 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
 309   CAPMAP(HF__FLAG)
 310 #undef HF__FLAG
 311   HF__END
 312 };
 313
 314 /* Build a structure in which we can capture the interesting data from the
 315  * auxiliary vector.
 316  */
 317 #define AUXUTYPE_i long
 318 #define AUXUTYPE_u unsigned long
 319 #define AUXUTYPE_p const void *
 320 struct auxprobe {
 321 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
 322   WANTAUX(AUXPROBE__SLOT)
 323 #undef AUXPROBE_SLOT
 324 };
 325
 326 /* --- @probe_hwcaps@ --- *
 327  *
 328  * Arguments:   ---
 329  *
 330  * Returns:     ---
 331  *
 332  * Use:         Attempt to find the auxiliary vector (which is well hidden)
 333  *              and discover interesting features from it.
 334  */
 335
 336 static void probe_hwcaps(void)
 337 {
 338   unsigned hw = HF_PROBED;
 339   struct auxprobe probed = { 0 };
 340
 341   /* Populate `probed' with the information we manage to retrieve from the
 342    * auxiliary vector.  Slots we couldn't find are left zero-valued.
 343    */
 344 #if defined(HAVE_GETAUXVAL)
 345   /* Shiny new libc lets us request individual entry types.  This is almost
 346    * too easy.
 347    */
 348 #  define CAP__GET(type, ubranch, slot)                                 \
 349         probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
 350   WANTAUX(CAP__GET)
 351 #else
 352   /* Otherwise we're a bit stuck, really.  Modern Linux kernels make a copy
 353    * of the vector available in `/procc' so we could try that.
 354    *
 355    * The usual place is stuck on the end of the environment vector, but that
 356    * may well have moved, and we have no way of telling whether it has or
 357    * whether there was ever an auxiliary vector there at all; so don't do
 358    * that.
 359    */
 360   {
 361     FILE *fp = 0;
 362     unsigned char *p = 0, *q = 0;
 363     const struct auxentry *a;
 364     size_t sz, off, n;
 365
 366     /* Open the file and read it into a memory chunk. */
 367     if ((fp = fopen("/proc/self/auxv", "rb")) == 0) goto clean;
 368     sz = 4096; off = 0;
 369     if ((p = malloc(sz)) == 0) goto clean;
 370     for (;;) {
 371       n = fread(p + off, 1, sz - off, fp);
 372       off += n;
 373       if (off < sz) break;
 374       sz *= 2; if ((q = realloc(p, sz)) == 0) break;
 375       p = q;
 376     }
 377
 378     /* Work through the vector (or as much of it as we found) and extract the
 379      * types we're interested in.
 380      */
 381     for (a = (const struct auxentry *)p,
 382            n = sz/sizeof(struct auxentry);
 383          n--; a++) {
 384       switch (a->type) {
 385 #define CAP__SWITCH(type, ubranch, slot)                                \
 386         case type: probed.slot = a->value.ubranch; break;
 387         WANTAUX(CAP__SWITCH)
 388         case AT_NULL: goto clean;
 389       }
 390     }
 391
 392   clean:
 393     if (p) free(p);
 394     if (fp) fclose(fp);
 395   }
 396 #endif
 397
 398   /* Each CPU family now has to pick through what was found and stashed in
 399    * `probed', and set the appropriate flag bits in `hw'.
 400    */
 401 #if CPUFAM_ARMEL
 402   if (probed.hwcap & HWCAP_VFPv3) hw |= HF_ARM_VFP;
 403   if (probed.hwcap & HWCAP_NEON) hw |= HF_ARM_NEON;
 404   if (probed.hwcap & HWCAP_VFPD32) hw |= HF_ARM_D32;
 405   if (probed.hwcap & HWCAP_VFPv4) hw |= HF_ARM_V4;
 406 #  ifdef HWCAP2_AES
 407   if (probed.hwcap2 & HWCAP2_AES) hw |= HF_ARM_AES;
 408 #  endif
 409 #  ifdef HWCAP2_PMULL
 410   if (probed.hwcap2 & HWCAP2_PMULL) hw |= HF_ARM_PMULL;
 411 #  endif
 412 #endif
 413 #if CPUFAM_ARM64
 414   if (probed.hwcap & HWCAP_AES) hw |= HF_ARM_AES;
 415   if (probed.hwcap & HWCAP_PMULL) hw |= HF_ARM_PMULL;
 416 #endif
 417
 418   /* Store the bitmask of features we probed for everyone to see. */
 419   DISPATCH_STORE(hwcaps, hw);
 420
 421   /* Finally, make a report about the things we found.  (Doing this earlier
 422    * will pointlessly widen the window in which multiple threads will do the
 423    * above auxiliary-vector probing.)
 424    */
 425 #define CAP__DEBUG(feat, tok)                                           \
 426   dispatch_debug("check auxv for feature `%s': %s", tok,                \
 427                  hw & HF_##feat ? "available" : "absent");
 428   CAPMAP(CAP__DEBUG)
 429 #undef CAP__DEBUG
 430 }
 431
 432 /* --- @get_hwcaps@ --- *
 433  *
 434  * Arguments:   ---
 435  *
 436  * Returns:     A mask of hardware capabilities and other features, as probed
 437  *              from the auxiliary vector.
 438  */
 439
 440 static unsigned get_hwcaps(void)
 441 {
 442   unsigned hw;
 443
 444   DISPATCH_LOAD(hwcaps, hw);
 445   if (!(hwcaps & HF_PROBED)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps, hw); }
 446   return (hw);
 447 }
 448
 449 #endif
 450
 451 /*----- External interface ------------------------------------------------*/
 452
 453 /* --- @dispatch_debug@ --- *
 454  *
 455  * Arguments:   @const char *fmt@ = a format string
 456  *              @...@ = additional arguments
 457  *
 458  * Returns:     ---
 459  *
 460  * Use:         Writes a formatted message to standard output if dispatch
 461  *              debugging is enabled.
 462  */
 463
 464 void dispatch_debug(const char *fmt, ...)
 465 {
 466   va_list ap;
 467   const char *e = getenv("CATACOMB_CPUDISPATCH_DEBUG");
 468
 469   if (e && *e != 'n' && *e != '0') {
 470     va_start(ap, fmt);
 471     fputs("Catacomb CPUDISPATCH: ", stderr);
 472     vfprintf(stderr, fmt, ap);
 473     fputc('\n', stderr);
 474     va_end(ap);
 475   }
 476 }
 477
 478 /* --- @check_env@ --- *
 479  *
 480  * Arguments:   @const char *ftok@ = feature token
 481  *
 482  * Returns:     Zero if the feature is forced off; positive if it's forced
 483  *              on; negative if the user hasn't decided.
 484  *
 485  * Use:         Checks the environment variable `CATACOMB_CPUFEAT' for the
 486  *              feature token @ftok@.  The variable, if it exists, should be
 487  *              a space-separated sequence of `+tok' and `-tok' items.  These
 488  *              tokens may end in `*', which matches any suffix.
 489  */
 490
 491 static int IGNORABLE check_env(const char *ftok)
 492 {
 493   const char *p, *q, *pp;
 494   int d;
 495
 496   p = getenv("CATACOMB_CPUFEAT");
 497   if (!p) return (-1);
 498
 499   for (;;) {
 500     while (isspace((unsigned char)*p)) p++;
 501     if (!*p) return (-1);
 502     switch (*p) {
 503       case '+': d = +1; p++; break;
 504       case '-': d =  0; p++; break;
 505       default:  d = -1;      break;
 506     }
 507     for (q = p; *q && !isspace((unsigned char)*q); q++);
 508     if (d >= 0) {
 509       for (pp = ftok; p < q && *pp && *p == *pp; p++, pp++);
 510       if ((p == q && !*pp) || (*p == '*' && p + 1 == q)) return (d);
 511     }
 512     p = q;
 513   }
 514   return (-1);
 515 }
 516
 517 /* --- @cpu_feature_p@ --- *
 518  *
 519  * Arguments:   @unsigned feat@ = a @CPUFEAT_...@ code
 520  *
 521  * Returns:     Nonzero if the feature is available.
 522  */
 523
 524 #include <stdio.h>
 525
 526 static int IGNORABLE
 527   feat_debug(const char *ftok, const char *check, int verdict)
 528 {
 529   if (verdict >= 0) {
 530     dispatch_debug("feature `%s': %s -> %s", ftok, check,
 531                    verdict ? "available" : "absent");
 532   }
 533   return (verdict);
 534 }
 535
 536 int cpu_feature_p(int feat)
 537 {
 538   int IGNORABLE f;
 539   IGNORE(f);
 540 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat:             \
 541   if ((f = feat_debug(ftok, "environment override",                     \
 542                       check_env(ftok))) >= 0)                           \
 543     return (f);                                                         \
 544   else                                                                  \
 545     return (feat_debug(ftok, "runtime probe", cond));
 546
 547   switch (feat) {
 548 #if CPUFAM_X86 || CPUFAM_AMD64
 549     CASE_CPUFEAT(X86_SSE2, "x86:sse2",
 550                  cpuid_features_p(CPUID1D_SSE2, 0) &&
 551                  xmm_registers_available_p());
 552     CASE_CPUFEAT(X86_AESNI, "x86:aesni",
 553                  cpuid_features_p(CPUID1D_SSE2, CPUID1C_AESNI) &&
 554                  xmm_registers_available_p());
 555     CASE_CPUFEAT(X86_RDRAND, "x86:rdrand",
 556                  cpuid_features_p(0, CPUID1C_RDRAND));
 557     CASE_CPUFEAT(X86_AVX, "x86:avx",
 558                  cpuid_features_p(0, CPUID1C_AVX) &&
 559                  xmm_registers_available_p());
 560     CASE_CPUFEAT(X86_SSSE3, "x86:ssse3",
 561                  cpuid_features_p(0, CPUID1C_SSSE3) &&
 562                  xmm_registers_available_p());
 563     CASE_CPUFEAT(X86_PCLMUL, "x86:pclmul",
 564                  cpuid_features_p(0, CPUID1C_PCLMUL) &&
 565                  xmm_registers_available_p());
 566 #endif
 567 #ifdef CAPMAP
 568 #  define FEATP__CASE(feat, tok)                                        \
 569         CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
 570     CAPMAP(FEATP__CASE)
 571 #undef FEATP__CASE
 572 #endif
 573     default:
 574       dispatch_debug("denying unknown feature %d", feat);
 575       return (0);
 576   }
 577 #undef CASE_CPUFEAT
 578 }
 579
 580 /*----- That's all, folks -------------------------------------------------*/