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