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_AESNI (1u << 25)
  50 #  define CPUID1C_RDRAND (1u << 30)
  51
  52 struct cpuid { unsigned a, b, c, d; };
  53
  54 /* --- @cpuid@ --- *
  55  *
  56  * Arguments:   @struct cpuid *cc@ = where to write the result
  57  *              @unsigned a, c@ = EAX and ECX registers to set
  58  *
  59  * Returns:     ---
  60  *
  61  * Use:         Minimal C wrapper around the x86 `CPUID' instruction.  Checks
  62  *              that the instruction is actually available before invoking
  63  *              it; fills the output structure with zero if it's not going to
  64  *              work.
  65  */
  66
  67 #ifdef __GNUC__
  68 #  if CPUFAM_X86
  69 static __inline__ unsigned getflags(void)
  70   { unsigned f; __asm__ ("pushf; popl %0" : "=g" (f)); return (f); }
  71 static __inline__ unsigned setflags(unsigned f)
  72 {
  73   unsigned ff;
  74   __asm__ ("pushf; pushl %1; popf; pushf; popl %0; popf"
  75            : "=r" (ff)
  76            : "r" (f));
  77   return (ff);
  78 }
  79 #  else
  80 static __inline__ unsigned long getflags(void)
  81   { unsigned long f; __asm__ ("pushf; popq %0" : "=g" (f)); return (f); }
  82 static __inline__ unsigned long long setflags(unsigned long f)
  83 {
  84   unsigned long ff;
  85   __asm__ ("pushf; pushq %1; popf; pushf; popq %0; popf"
  86            : "=r" (ff)
  87            : "r" (f));
  88   return (ff);
  89 }
  90 #  endif
  91 #endif
  92
  93 static void cpuid(struct cpuid *cc, unsigned a, unsigned c)
  94 {
  95 #ifdef __GNUC__
  96   unsigned f;
  97 #endif
  98
  99   cc->a = cc->b = cc->c = cc->d = 0;
 100
 101 #ifdef __GNUC__
 102   /* Stupid dance to detect whether the CPUID instruction is available. */
 103   f = getflags();
 104   if (!(setflags(f |  EFLAGS_ID) & EFLAGS_ID) ||
 105         setflags(f & ~EFLAGS_ID) & EFLAGS_ID) {
 106     dispatch_debug("CPUID instruction not available");
 107     return;
 108   }
 109   setflags(f);
 110
 111   /* Alas, EBX is magical in PIC code, so abuse ESI instead.  This isn't
 112    * pretty, but it works.
 113    */
 114 #  if CPUFAM_X86
 115   __asm__ ("pushl %%ebx; cpuid; movl %%ebx, %%esi; popl %%ebx"
 116            : "=a" (cc->a), "=S" (cc->b), "=c" (cc->c), "=d" (cc->d)
 117            : "a" (a) , "c" (c));
 118 #  elif CPUFAM_AMD64
 119   __asm__ ("pushq %%rbx; cpuid; movl %%ebx, %%esi; popq %%rbx"
 120            : "=a" (cc->a), "=S" (cc->b), "=c" (cc->c), "=d" (cc->d)
 121            : "a" (a) , "c" (c));
 122 #  else
 123 #    error "I'm confused."
 124 #  endif
 125   dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x",
 126                  a, c, cc->a, cc->b, cc->c, cc->d);
 127 #else
 128   dispatch_debug("GNU inline assembler not available; can't CPUID");
 129 #endif
 130 }
 131
 132 static unsigned cpuid_maxleaf(void)
 133   { struct cpuid c; cpuid(&c, 0, 0); return (c.a); }
 134
 135 /* --- @cpuid_features_p@ --- *
 136  *
 137  * Arguments:   @unsigned dbits@ = bits to check in EDX
 138  *              @unsigned cbits@ = bits to check in ECX
 139  *
 140  * Returns:     Nonzero if all the requested bits are set in the CPUID result
 141  *              on leaf 1.
 142  */
 143
 144 static int cpuid_features_p(unsigned dbits, unsigned cbits)
 145 {
 146   struct cpuid c;
 147   if (cpuid_maxleaf() < 1) return (0);
 148   cpuid(&c, 1, 0);
 149   return ((c.d & dbits) == dbits && (c.c & cbits) == cbits);
 150 }
 151
 152 /* --- @xmm_registers_available_p@ --- *
 153  *
 154  * Arguments:   ---
 155  *
 156  * Returns:     Nonzero if the operating system has made the XMM registers
 157  *              available for use.
 158  */
 159
 160 static int xmm_registers_available_p(void)
 161 {
 162 #ifdef __GNUC__
 163   unsigned f;
 164   /* This hack is by Agner Fog.  Use FXSAVE/FXRSTOR to figure out whether the
 165    * XMM registers are actually alive.
 166    */
 167   if (!cpuid_features_p(CPUID1D_FXSR, 0)) return (0);
 168 #  if CPUFAM_X86
 169   __asm__ ("movl %%esp, %%edx; subl $512, %%esp; andl $~15, %%esp\n"
 170            "fxsave (%%esp)\n"
 171            "movl 160(%%esp), %%eax; xorl $0xaaaa5555, 160(%%esp)\n"
 172            "fxrstor (%%esp); fxsave (%%esp)\n"
 173            "movl 160(%%esp), %%ecx; movl %%eax, 160(%%esp)\n"
 174            "fxrstor (%%esp); movl %%edx, %%esp\n"
 175            "xorl %%ecx, %%eax"
 176            : "=a" (f)
 177            : /* no inputs */
 178            : "%ecx", "%edx");
 179 #  elif CPUFAM_AMD64
 180   __asm__ ("movq %%rsp, %%rdx; subq $512, %%rsp; andq $~15, %%rsp\n"
 181            "fxsave (%%rsp)\n"
 182            "movl 160(%%rsp), %%eax; xorl $0xaaaa5555, 160(%%rsp)\n"
 183            "fxrstor (%%rsp); fxsave (%%rsp)\n"
 184            "movl 160(%%rsp), %%ecx; movl %%eax, 160(%%rsp)\n"
 185            "fxrstor (%%rsp); movq %%rdx, %%rsp\n"
 186            "xorl %%ecx, %%eax"
 187            : "=a" (f)
 188            : /* no inputs */
 189            : "%ecx", "%rdx");
 190 #  else
 191 #    error "I'm confused."
 192 #  endif
 193   dispatch_debug("XMM registers %savailable", f ? "" : "not ");
 194   return (f);
 195 #else
 196   dispatch_debug("GNU inline assembler not available; can't check for XMM");
 197   return (0);
 198 #endif
 199 }
 200
 201 /* --- @rdrand_works_p@ --- *
 202  *
 203  *
 204  * Arguments:   ---
 205  *
 206  * Returns:     Nonzero if the `rdrand' instruction actually works.  Assumes
 207  *              that it's already been verified to be safe to issue.
 208  */
 209
 210 #ifdef __GNUC__
 211 static int rdrand(unsigned *x)
 212 {
 213   int i, rc;
 214   unsigned _t;
 215
 216   i = 16;
 217   __asm__ ("" : "=g" (_t));
 218   __asm__ ("0: rdrand %2; jc 1f; decl %1; jnz 0b\n"
 219            "mov $-1, %0; jmp 9f\n"
 220            "1: movl %2, (%3); xorl %0, %0\n"
 221            "9:"
 222            : "=r" (rc), "+r" (i), "+r" (_t)
 223            : "r" (x)
 224            : "cc");
 225   return (rc);
 226 }
 227 #endif
 228
 229 static int rdrand_works_p(void)
 230 {
 231   unsigned ref, x, i;
 232
 233   /* Check that it doesn't always give the same answer.  Try four times: this
 234    * will fail with probability %$2^{-128}$% with a truly random generator,
 235    * which seems fair enough.
 236    */
 237   if (rdrand(&ref)) goto fail;
 238   for (i = 0; i < 4; i++) {
 239     if (rdrand(&x)) goto fail;
 240     if (x != ref) goto not_stuck;
 241   }
 242   dispatch_debug("RDRAND always returns 0x%08x!", ref);
 243   return (0);
 244
 245 not_stuck:
 246   dispatch_debug("RDRAND instruction looks plausible");
 247   return (1);
 248
 249 fail:
 250   dispatch_debug("RDRAND instruction fails too often");
 251   return (0);
 252 }
 253
 254 #endif
 255
 256 /*----- General feature probing using auxiliary vectors -------------------*/
 257
 258 /* Try to find the system's definitions for auxiliary vector entries. */
 259 #ifdef HAVE_SYS_AUXV_H
 260 #  include <sys/auxv.h>
 261 #endif
 262 #ifdef HAVE_LINUX_AUXVEC_H
 263 #  include <linux/auxvec.h>
 264 #endif
 265 #ifdef HAVE_ASM_HWCAP_H
 266 #  include <asm/hwcap.h>
 267 #endif
 268
 269 /* The type of entries in the auxiliary vector.  I'm assuming that `unsigned
 270  * long' matches each platform's word length; if this is false then we'll
 271  * need some host-specific tweaking here.
 272  */
 273 union auxval { long i; unsigned long u; const void *p; };
 274 struct auxentry { unsigned long type; union auxval value; };
 275
 276 /* Register each CPU family's interest in the auxiliary vector.  Make sure
 277  * that the necessary entry types are defined.  This is primarily ordered by
 278  * entry type to minimize duplication.
 279  */
 280 #if defined(AT_HWCAP) && CPUFAM_ARMEL
 281 #  define WANT_ANY 1
 282 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 283 #endif
 284
 285 #if defined(AT_HWCAP) && CPUFAM_ARM64
 286 #  define WANT_ANY 1
 287 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 288 #endif
 289
 290 #if defined(AT_HWCAP2) && CPUFAM_ARMEL
 291 #  define WANT_ANY 1
 292 #  define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
 293 #endif
 294
 295 /* If we couldn't find any interesting entries then we can switch all of this
 296  * machinery off.  Also do that if we have no means for atomic updates.
 297  */
 298 #if WANT_ANY && CPU_DISPATCH_P
 299
 300 /* The main output of this section is a bitmask of detected features.  The
 301  * least significant bit will be set if we've tried to probe.  Always access
 302  * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
 303  */
 304 static unsigned hwcaps = 0;
 305
 306 /* For each potentially interesting type which turned out not to exist or be
 307  * wanted, define a dummy macro for the sake of the next step.
 308  */
 309 #ifndef WANT_AT_HWCAP
 310 #  define WANT_AT_HWCAP(_)
 311 #endif
 312 #ifndef WANT_AT_HWCAP2
 313 #  define WANT_AT_HWCAP2(_)
 314 #endif
 315
 316 /* For each CPU family, define two lists.
 317  *
 318  *   * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
 319  *     family tried to register interest in above.  Each entry contains the
 320  *     interesting auxiliary vector entry type, the name of the union branch
 321  *     for its value, and the name of the slot in `struct auxprobe' in which
 322  *     to store the value.
 323  *
 324  *   * `CAPMAP' is a list describing the output features which the CPU family
 325  *     intends to satisfy from the auxiliary vector.  Each entry contains a
 326  *     feature name suffix, and the token name (for `check_env').
 327  */
 328 #if CPUFAM_ARMEL
 329 #  define WANTAUX(_)                                                    \
 330         WANT_AT_HWCAP(_)                                                \
 331         WANT_AT_HWCAP2(_)
 332 #  define CAPMAP(_)                                                     \
 333         _(ARM_VFP, "arm:vfp")                                           \
 334         _(ARM_NEON, "arm:neon")                                         \
 335         _(ARM_V4, "arm:v4")                                             \
 336         _(ARM_D32, "arm:d32")                                           \
 337         _(ARM_AES, "arm:aes")
 338 #endif
 339 #if CPUFAM_ARM64
 340 #  define WANTAUX(_)                                                    \
 341         WANT_AT_HWCAP(_)
 342 #  define CAPMAP(_)                                                     \
 343         _(ARM_NEON, "arm:neon")                                         \
 344         _(ARM_AES, "arm:aes")
 345 #endif
 346
 347 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
 348 enum {
 349   HFI_PROBED = 0,
 350 #define HFI__ENUM(feat, tok) HFI_##feat,
 351   CAPMAP(HFI__ENUM)
 352 #undef HFI__ENUM
 353   HFI__END
 354 };
 355 enum {
 356   HF_PROBED = 1,
 357 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
 358   CAPMAP(HF__FLAG)
 359 #undef HF__FLAG
 360   HF__END
 361 };
 362
 363 /* Build a structure in which we can capture the interesting data from the
 364  * auxiliary vector.
 365  */
 366 #define AUXUTYPE_i long
 367 #define AUXUTYPE_u unsigned long
 368 #define AUXUTYPE_p const void *
 369 struct auxprobe {
 370 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
 371   WANTAUX(AUXPROBE__SLOT)
 372 #undef AUXPROBE_SLOT
 373 };
 374
 375 /* --- @probe_hwcaps@ --- *
 376  *
 377  * Arguments:   ---
 378  *
 379  * Returns:     ---
 380  *
 381  * Use:         Attempt to find the auxiliary vector (which is well hidden)
 382  *              and discover interesting features from it.
 383  */
 384
 385 static void probe_hwcaps(void)
 386 {
 387   unsigned hw = HF_PROBED;
 388   struct auxprobe probed = { 0 };
 389
 390   /* Populate `probed' with the information we manage to retrieve from the
 391    * auxiliary vector.  Slots we couldn't find are left zero-valued.
 392    */
 393 #if defined(HAVE_GETAUXVAL)
 394   /* Shiny new libc lets us request individual entry types.  This is almost
 395    * too easy.
 396    */
 397 #  define CAP__GET(type, ubranch, slot)                                 \
 398         probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
 399   WANTAUX(CAP__GET)
 400 #else
 401   /* Otherwise we're a bit stuck, really.  Modern Linux kernels make a copy
 402    * of the vector available in `/procc' so we could try that.
 403    *
 404    * The usual place is stuck on the end of the environment vector, but that
 405    * may well have moved, and we have no way of telling whether it has or
 406    * whether there was ever an auxiliary vector there at all; so don't do
 407    * that.
 408    */
 409   {
 410     FILE *fp = 0;
 411     unsigned char *p = 0, *q = 0;
 412     const struct auxentry *a;
 413     size_t sz, off, n;
 414
 415     /* Open the file and read it into a memory chunk. */
 416     if ((fp = fopen("/proc/self/auxv", "rb")) == 0) goto clean;
 417     sz = 4096; off = 0;
 418     if ((p = malloc(sz)) == 0) goto clean;
 419     for (;;) {
 420       n = fread(p + off, 1, sz - off, fp);
 421       off += n;
 422       if (off < sz) break;
 423       sz *= 2; if ((q = realloc(p, sz)) == 0) break;
 424       p = q;
 425     }
 426
 427     /* Work through the vector (or as much of it as we found) and extract the
 428      * types we're interested in.
 429      */
 430     for (a = (const struct auxentry *)p,
 431            n = sz/sizeof(struct auxentry);
 432          n--; a++) {
 433       switch (a->type) {
 434 #define CAP__SWITCH(type, ubranch, slot)                                \
 435         case type: probed.slot = a->value.ubranch; break;
 436         WANTAUX(CAP__SWITCH)
 437         case AT_NULL: goto clean;
 438       }
 439     }
 440
 441   clean:
 442     if (p) free(p);
 443     if (fp) fclose(fp);
 444   }
 445 #endif
 446
 447   /* Each CPU family now has to pick through what was found and stashed in
 448    * `probed', and set the appropriate flag bits in `hw'.
 449    */
 450 #if CPUFAM_ARMEL
 451   if (probed.hwcap & HWCAP_VFPv3) hw |= HF_ARM_VFP;
 452   if (probed.hwcap & HWCAP_NEON) hw |= HF_ARM_NEON;
 453   if (probed.hwcap & HWCAP_VFPD32) hw |= HF_ARM_D32;
 454   if (probed.hwcap & HWCAP_VFPv4) hw |= HF_ARM_V4;
 455 #  ifdef HWCAP2_AES
 456   if (probed.hwcap2 & HWCAP2_AES) hw |= HF_ARM_AES;
 457 #  endif
 458 #endif
 459 #if CPUFAM_ARM64
 460   if (probed.hwcap & HWCAP_ASIMD) hw |= HF_ARM_NEON;
 461   if (probed.hwcap & HWCAP_AES) hw |= HF_ARM_AES;
 462 #endif
 463
 464   /* Store the bitmask of features we probed for everyone to see. */
 465   DISPATCH_STORE(hwcaps, hw);
 466
 467   /* Finally, make a report about the things we found.  (Doing this earlier
 468    * will pointlessly widen the window in which multiple threads will do the
 469    * above auxiliary-vector probing.)
 470    */
 471 #define CAP__DEBUG(feat, tok)                                           \
 472   dispatch_debug("check auxv for feature `%s': %s", tok,                \
 473                  hw & HF_##feat ? "available" : "absent");
 474   CAPMAP(CAP__DEBUG)
 475 #undef CAP__DEBUG
 476 }
 477
 478 /* --- @get_hwcaps@ --- *
 479  *
 480  * Arguments:   ---
 481  *
 482  * Returns:     A mask of hardware capabilities and other features, as probed
 483  *              from the auxiliary vector.
 484  */
 485
 486 static unsigned get_hwcaps(void)
 487 {
 488   unsigned hw;
 489
 490   DISPATCH_LOAD(hwcaps, hw);
 491   if (!(hwcaps & HF_PROBED)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps, hw); }
 492   return (hw);
 493 }
 494
 495 #endif
 496
 497 /*----- External interface ------------------------------------------------*/
 498
 499 /* --- @dispatch_debug@ --- *
 500  *
 501  * Arguments:   @const char *fmt@ = a format string
 502  *              @...@ = additional arguments
 503  *
 504  * Returns:     ---
 505  *
 506  * Use:         Writes a formatted message to standard output if dispatch
 507  *              debugging is enabled.
 508  */
 509
 510 void dispatch_debug(const char *fmt, ...)
 511 {
 512   va_list ap;
 513   const char *e = getenv("CATACOMB_CPUDISPATCH_DEBUG");
 514
 515   if (e && *e != 'n' && *e != '0') {
 516     va_start(ap, fmt);
 517     fputs("Catacomb CPUDISPATCH: ", stderr);
 518     vfprintf(stderr, fmt, ap);
 519     fputc('\n', stderr);
 520     va_end(ap);
 521   }
 522 }
 523
 524 /* --- @check_env@ --- *
 525  *
 526  * Arguments:   @const char *ftok@ = feature token
 527  *
 528  * Returns:     Zero if the feature is forced off; positive if it's forced
 529  *              on; negative if the user hasn't decided.
 530  *
 531  * Use:         Checks the environment variable `CATACOMB_CPUFEAT' for the
 532  *              feature token @ftok@.  The variable, if it exists, should be
 533  *              a space-separated sequence of `+tok' and `-tok' items.  These
 534  *              tokens may end in `*', which matches any suffix.
 535  */
 536
 537 static int IGNORABLE check_env(const char *ftok)
 538 {
 539   const char *p, *q, *pp;
 540   int d;
 541
 542   p = getenv("CATACOMB_CPUFEAT");
 543   if (!p) return (-1);
 544
 545   for (;;) {
 546     while (isspace((unsigned char)*p)) p++;
 547     if (!*p) return (-1);
 548     switch (*p) {
 549       case '+': d = +1; p++; break;
 550       case '-': d =  0; p++; break;
 551       default:  d = -1;      break;
 552     }
 553     for (q = p; *q && !isspace((unsigned char)*q); q++);
 554     if (d >= 0) {
 555       for (pp = ftok; p < q && *pp && *p == *pp; p++, pp++);
 556       if ((p == q && !*pp) || (*p == '*' && p + 1 == q)) return (d);
 557     }
 558     p = q;
 559   }
 560   return (-1);
 561 }
 562
 563 /* --- @cpu_feature_p@ --- *
 564  *
 565  * Arguments:   @unsigned feat@ = a @CPUFEAT_...@ code
 566  *
 567  * Returns:     Nonzero if the feature is available.
 568  */
 569
 570 #include <stdio.h>
 571
 572 static int IGNORABLE
 573   feat_debug(const char *ftok, const char *check, int verdict)
 574 {
 575   if (verdict >= 0) {
 576     dispatch_debug("feature `%s': %s -> %s", ftok, check,
 577                    verdict ? "available" : "absent");
 578   }
 579   return (verdict);
 580 }
 581
 582 int cpu_feature_p(int feat)
 583 {
 584   int IGNORABLE f;
 585   IGNORE(f);
 586 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat:             \
 587   if ((f = feat_debug(ftok, "environment override",                     \
 588                       check_env(ftok))) >= 0)                           \
 589     return (f);                                                         \
 590   else                                                                  \
 591     return (feat_debug(ftok, "runtime probe", cond));
 592
 593   switch (feat) {
 594 #if CPUFAM_X86 || CPUFAM_AMD64
 595     CASE_CPUFEAT(X86_SSE2, "x86:sse2",
 596                  cpuid_features_p(CPUID1D_SSE2, 0) &&
 597                  xmm_registers_available_p());
 598     CASE_CPUFEAT(X86_AESNI, "x86:aesni",
 599                  cpuid_features_p(CPUID1D_SSE2, CPUID1C_AESNI) &&
 600                  xmm_registers_available_p());
 601     CASE_CPUFEAT(X86_RDRAND, "x86:rdrand",
 602                  cpuid_features_p(0, CPUID1C_RDRAND) && rdrand_works_p());
 603 #endif
 604 #ifdef CAPMAP
 605 #  define FEATP__CASE(feat, tok)                                        \
 606         CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
 607     CAPMAP(FEATP__CASE)
 608 #undef FEATP__CASE
 609 #endif
 610     default:
 611       dispatch_debug("denying unknown feature %d", feat);
 612       return (0);
 613   }
 614 #undef CASE_CPUFEAT
 615 }
 616
 617 /*----- That's all, folks -------------------------------------------------*/