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