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 #endif
 202
 203 /*----- General feature probing using auxiliary vectors -------------------*/
 204
 205 /* Try to find the system's definitions for auxiliary vector entries. */
 206 #ifdef HAVE_SYS_AUXV_H
 207 #  include <sys/auxv.h>
 208 #endif
 209 #ifdef HAVE_LINUX_AUXVEC_H
 210 #  include <linux/auxvec.h>
 211 #endif
 212 #ifdef HAVE_ASM_HWCAP_H
 213 #  include <asm/hwcap.h>
 214 #endif
 215
 216 /* The type of entries in the auxiliary vector.  I'm assuming that `unsigned
 217  * long' matches each platform's word length; if this is false then we'll
 218  * need some host-specific tweaking here.
 219  */
 220 union auxval { long i; unsigned long u; const void *p; };
 221 struct auxentry { unsigned long type; union auxval value; };
 222
 223 /* Register each CPU family's interest in the auxiliary vector.  Make sure
 224  * that the necessary entry types are defined.  This is primarily ordered by
 225  * entry type to minimize duplication.
 226  */
 227 #if defined(AT_HWCAP) && CPUFAM_ARMEL
 228 #  define WANT_ANY 1
 229 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 230 #endif
 231
 232 #if defined(AT_HWCAP) && CPUFAM_ARM64
 233 #  define WANT_ANY 1
 234 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 235 #endif
 236
 237 #if defined(AT_HWCAP2) && CPUFAM_ARMEL
 238 #  define WANT_ANY 1
 239 #  define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
 240 #endif
 241
 242 /* If we couldn't find any interesting entries then we can switch all of this
 243  * machinery off.  Also do that if we have no means for atomic updates.
 244  */
 245 #if WANT_ANY && CPU_DISPATCH_P
 246
 247 /* The main output of this section is a bitmask of detected features.  The
 248  * least significant bit will be set if we've tried to probe.  Always access
 249  * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
 250  */
 251 static unsigned hwcaps = 0;
 252
 253 /* For each potentially interesting type which turned out not to exist or be
 254  * wanted, define a dummy macro for the sake of the next step.
 255  */
 256 #ifndef WANT_AT_HWCAP
 257 #  define WANT_AT_HWCAP(_)
 258 #endif
 259 #ifndef WANT_AT_HWCAP2
 260 #  define WANT_AT_HWCAP2(_)
 261 #endif
 262
 263 /* For each CPU family, define two lists.
 264  *
 265  *   * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
 266  *     family tried to register interest in above.  Each entry contains the
 267  *     interesting auxiliary vector entry type, the name of the union branch
 268  *     for its value, and the name of the slot in `struct auxprobe' in which
 269  *     to store the value.
 270  *
 271  *   * `CAPMAP' is a list describing the output features which the CPU family
 272  *     intends to satisfy from the auxiliary vector.  Each entry contains a
 273  *     feature name suffix, and the token name (for `check_env').
 274  */
 275 #if CPUFAM_ARMEL
 276 #  define WANTAUX(_)                                                    \
 277         WANT_AT_HWCAP(_)                                                \
 278         WANT_AT_HWCAP2(_)
 279 #  define CAPMAP(_)                                                     \
 280         _(ARM_VFP, "arm:vfp")                                           \
 281         _(ARM_NEON, "arm:neon")                                         \
 282         _(ARM_V4, "arm:v4")                                             \
 283         _(ARM_D32, "arm:d32")                                           \
 284         _(ARM_AES, "arm:aes")
 285 #endif
 286 #if CPUFAM_ARM64
 287 #  define WANTAUX(_)                                                    \
 288         WANT_AT_HWCAP(_)
 289 #  define CAPMAP(_)                                                     \
 290         _(ARM_AES, "arm:aes")
 291 #endif
 292
 293 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
 294 enum {
 295   HFI_PROBED = 0,
 296 #define HFI__ENUM(feat, tok) HFI_##feat,
 297   CAPMAP(HFI__ENUM)
 298 #undef HFI__ENUM
 299   HFI__END
 300 };
 301 enum {
 302   HF_PROBED = 1,
 303 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
 304   CAPMAP(HF__FLAG)
 305 #undef HF__FLAG
 306   HF__END
 307 };
 308
 309 /* Build a structure in which we can capture the interesting data from the
 310  * auxiliary vector.
 311  */
 312 #define AUXUTYPE_i long
 313 #define AUXUTYPE_u unsigned long
 314 #define AUXUTYPE_p const void *
 315 struct auxprobe {
 316 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
 317   WANTAUX(AUXPROBE__SLOT)
 318 #undef AUXPROBE_SLOT
 319 };
 320
 321 /* --- @probe_hwcaps@ --- *
 322  *
 323  * Arguments:   ---
 324  *
 325  * Returns:     ---
 326  *
 327  * Use:         Attempt to find the auxiliary vector (which is well hidden)
 328  *              and discover interesting features from it.
 329  */
 330
 331 static void probe_hwcaps(void)
 332 {
 333   unsigned hw = HF_PROBED;
 334   struct auxprobe probed = { 0 };
 335
 336   /* Populate `probed' with the information we manage to retrieve from the
 337    * auxiliary vector.  Slots we couldn't find are left zero-valued.
 338    */
 339 #if defined(HAVE_GETAUXVAL)
 340   /* Shiny new libc lets us request individual entry types.  This is almost
 341    * too easy.
 342    */
 343 #  define CAP__GET(type, ubranch, slot)                                 \
 344         probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
 345   WANTAUX(CAP__GET)
 346 #else
 347   /* Otherwise we're a bit stuck, really.  Modern Linux kernels make a copy
 348    * of the vector available in `/procc' so we could try that.
 349    *
 350    * The usual place is stuck on the end of the environment vector, but that
 351    * may well have moved, and we have no way of telling whether it has or
 352    * whether there was ever an auxiliary vector there at all; so don't do
 353    * that.
 354    */
 355   {
 356     FILE *fp = 0;
 357     unsigned char *p = 0, *q = 0;
 358     const struct auxentry *a;
 359     size_t sz, off, n;
 360
 361     /* Open the file and read it into a memory chunk. */
 362     if ((fp = fopen("/proc/self/auxv", "rb")) == 0) goto clean;
 363     sz = 4096; off = 0;
 364     if ((p = malloc(sz)) == 0) goto clean;
 365     for (;;) {
 366       n = fread(p + off, 1, sz - off, fp);
 367       off += n;
 368       if (off < sz) break;
 369       sz *= 2; if ((q = realloc(p, sz)) == 0) break;
 370       p = q;
 371     }
 372
 373     /* Work through the vector (or as much of it as we found) and extract the
 374      * types we're interested in.
 375      */
 376     for (a = (const struct auxentry *)p,
 377            n = sz/sizeof(struct auxentry);
 378          n--; a++) {
 379       switch (a->type) {
 380 #define CAP__SWITCH(type, ubranch, slot)                                \
 381         case type: probed.slot = a->value.ubranch; break;
 382         WANTAUX(CAP__SWITCH)
 383         case AT_NULL: goto clean;
 384       }
 385     }
 386
 387   clean:
 388     if (p) free(p);
 389     if (fp) fclose(fp);
 390   }
 391 #endif
 392
 393   /* Each CPU family now has to pick through what was found and stashed in
 394    * `probed', and set the appropriate flag bits in `hw'.
 395    */
 396 #if CPUFAM_ARMEL
 397   if (probed.hwcap & HWCAP_VFPv3) hw |= HF_ARM_VFP;
 398   if (probed.hwcap & HWCAP_NEON) hw |= HF_ARM_NEON;
 399   if (probed.hwcap & HWCAP_VFPD32) hw |= HF_ARM_D32;
 400   if (probed.hwcap & HWCAP_VFPv4) hw |= HF_ARM_V4;
 401 #  ifdef HWCAP2_AES
 402   if (probed.hwcap2 & HWCAP2_AES) hw |= HF_ARM_AES;
 403 #  endif
 404 #endif
 405 #if CPUFAM_ARM64
 406   if (probed.hwcap & HWCAP_AES) hw |= HF_ARM_AES;
 407 #endif
 408
 409   /* Store the bitmask of features we probed for everyone to see. */
 410   DISPATCH_STORE(hwcaps, hw);
 411
 412   /* Finally, make a report about the things we found.  (Doing this earlier
 413    * will pointlessly widen the window in which multiple threads will do the
 414    * above auxiliary-vector probing.)
 415    */
 416 #define CAP__DEBUG(feat, tok)                                           \
 417   dispatch_debug("check auxv for feature `%s': %s", tok,                \
 418                  hw & HF_##feat ? "available" : "absent");
 419   CAPMAP(CAP__DEBUG)
 420 #undef CAP__DEBUG
 421 }
 422
 423 /* --- @get_hwcaps@ --- *
 424  *
 425  * Arguments:   ---
 426  *
 427  * Returns:     A mask of hardware capabilities and other features, as probed
 428  *              from the auxiliary vector.
 429  */
 430
 431 static unsigned get_hwcaps(void)
 432 {
 433   unsigned hw;
 434
 435   DISPATCH_LOAD(hwcaps, hw);
 436   if (!(hwcaps & HF_PROBED)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps, hw); }
 437   return (hw);
 438 }
 439
 440 #endif
 441
 442 /*----- External interface ------------------------------------------------*/
 443
 444 /* --- @dispatch_debug@ --- *
 445  *
 446  * Arguments:   @const char *fmt@ = a format string
 447  *              @...@ = additional arguments
 448  *
 449  * Returns:     ---
 450  *
 451  * Use:         Writes a formatted message to standard output if dispatch
 452  *              debugging is enabled.
 453  */
 454
 455 void dispatch_debug(const char *fmt, ...)
 456 {
 457   va_list ap;
 458   const char *e = getenv("CATACOMB_CPUDISPATCH_DEBUG");
 459
 460   if (e && *e != 'n' && *e != '0') {
 461     va_start(ap, fmt);
 462     fputs("Catacomb CPUDISPATCH: ", stderr);
 463     vfprintf(stderr, fmt, ap);
 464     fputc('\n', stderr);
 465     va_end(ap);
 466   }
 467 }
 468
 469 /* --- @check_env@ --- *
 470  *
 471  * Arguments:   @const char *ftok@ = feature token
 472  *
 473  * Returns:     Zero if the feature is forced off; positive if it's forced
 474  *              on; negative if the user hasn't decided.
 475  *
 476  * Use:         Checks the environment variable `CATACOMB_CPUFEAT' for the
 477  *              feature token @ftok@.  The variable, if it exists, should be
 478  *              a space-separated sequence of `+tok' and `-tok' items.  These
 479  *              tokens may end in `*', which matches any suffix.
 480  */
 481
 482 static int IGNORABLE check_env(const char *ftok)
 483 {
 484   const char *p, *q, *pp;
 485   int d;
 486
 487   p = getenv("CATACOMB_CPUFEAT");
 488   if (!p) return (-1);
 489
 490   for (;;) {
 491     while (isspace((unsigned char)*p)) p++;
 492     if (!*p) return (-1);
 493     switch (*p) {
 494       case '+': d = +1; p++; break;
 495       case '-': d =  0; p++; break;
 496       default:  d = -1;      break;
 497     }
 498     for (q = p; *q && !isspace((unsigned char)*q); q++);
 499     if (d >= 0) {
 500       for (pp = ftok; p < q && *pp && *p == *pp; p++, pp++);
 501       if ((p == q && !*pp) || (*p == '*' && p + 1 == q)) return (d);
 502     }
 503     p = q;
 504   }
 505   return (-1);
 506 }
 507
 508 /* --- @cpu_feature_p@ --- *
 509  *
 510  * Arguments:   @unsigned feat@ = a @CPUFEAT_...@ code
 511  *
 512  * Returns:     Nonzero if the feature is available.
 513  */
 514
 515 #include <stdio.h>
 516
 517 static int IGNORABLE
 518   feat_debug(const char *ftok, const char *check, int verdict)
 519 {
 520   if (verdict >= 0) {
 521     dispatch_debug("feature `%s': %s -> %s", ftok, check,
 522                    verdict ? "available" : "absent");
 523   }
 524   return (verdict);
 525 }
 526
 527 int cpu_feature_p(int feat)
 528 {
 529   int IGNORABLE f;
 530   IGNORE(f);
 531 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat:             \
 532   if ((f = feat_debug(ftok, "environment override",                     \
 533                       check_env(ftok))) >= 0)                           \
 534     return (f);                                                         \
 535   else                                                                  \
 536     return (feat_debug(ftok, "runtime probe", cond));
 537
 538   switch (feat) {
 539 #if CPUFAM_X86 || CPUFAM_AMD64
 540     CASE_CPUFEAT(X86_SSE2, "x86:sse2",
 541                  cpuid_features_p(CPUID1D_SSE2, 0) &&
 542                  xmm_registers_available_p());
 543     CASE_CPUFEAT(X86_AESNI, "x86:aesni",
 544                  cpuid_features_p(CPUID1D_SSE2, CPUID1C_AESNI) &&
 545                  xmm_registers_available_p());
 546     CASE_CPUFEAT(X86_RDRAND, "x86:rdrand",
 547                  cpuid_features_p(0, CPUID1C_RDRAND));
 548 #endif
 549 #ifdef CAPMAP
 550 #  define FEATP__CASE(feat, tok)                                        \
 551         CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
 552     CAPMAP(FEATP__CASE)
 553 #undef FEATP__CASE
 554 #endif
 555     default:
 556       dispatch_debug("denying unknown feature %d", feat);
 557       return (0);
 558   }
 559 #undef CASE_CPUFEAT
 560 }
 561
 562 /*----- That's all, folks -------------------------------------------------*/