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 CPUID1D_SSE2 (1u << 26)
  47 #  define CPUID1D_FXSR (1u << 24)
  48 #  define CPUID1C_PCLMUL (1u << 1)
  49 #  define CPUID1C_SSSE3 (1u << 9)
  50 #  define CPUID1C_AESNI (1u << 25)
  51 #  define CPUID1C_AVX (1u << 28)
  52 #  define CPUID1C_RDRAND (1u << 30)
  53
  54 struct cpuid { unsigned a, b, c, d; };
  55 extern int dispatch_x86ish_cpuid(struct cpuid *, unsigned a, unsigned c);
  56 extern int dispatch_x86ish_xmmregisters_p(void);
  57
  58 static void cpuid(struct cpuid *cc, unsigned a, unsigned c)
  59 {
  60   int rc = dispatch_x86ish_cpuid(cc, a, c);
  61   if (rc)
  62     dispatch_debug("CPUID instruction not available");
  63   else
  64     dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x",
  65                    a, c, cc->a, cc->b, cc->c, cc->d);
  66 }
  67
  68 static unsigned cpuid_maxleaf(void)
  69   { struct cpuid c; cpuid(&c, 0, 0); return (c.a); }
  70
  71 /* --- @cpuid_features_p@ --- *
  72  *
  73  * Arguments:   @unsigned dbits@ = bits to check in EDX
  74  *              @unsigned cbits@ = bits to check in ECX
  75  *
  76  * Returns:     Nonzero if all the requested bits are set in the CPUID result
  77  *              on leaf 1.
  78  */
  79
  80 static int cpuid_features_p(unsigned dbits, unsigned cbits)
  81 {
  82   struct cpuid c;
  83   if (cpuid_maxleaf() < 1) return (0);
  84   cpuid(&c, 1, 0);
  85   return ((c.d & dbits) == dbits && (c.c & cbits) == cbits);
  86 }
  87
  88 /* --- @xmm_registers_available_p@ --- *
  89  *
  90  * Arguments:   ---
  91  *
  92  * Returns:     Nonzero if the operating system has made the XMM registers
  93  *              available for use.
  94  */
  95
  96 static int xmm_registers_available_p(void)
  97 {
  98   int f = dispatch_x86ish_xmmregisters_p();
  99
 100   dispatch_debug("XMM registers %savailable", f ? "" : "not ");
 101   return (f);
 102 }
 103
 104 #endif
 105
 106 /*----- General feature probing using auxiliary vectors -------------------*/
 107
 108 /* Try to find the system's definitions for auxiliary vector entries. */
 109 #ifdef HAVE_SYS_AUXV_H
 110 #  include <sys/auxv.h>
 111 #endif
 112 #ifdef HAVE_LINUX_AUXVEC_H
 113 #  include <linux/auxvec.h>
 114 #endif
 115 #ifdef HAVE_ASM_HWCAP_H
 116 #  include <asm/hwcap.h>
 117 #endif
 118
 119 /* The type of entries in the auxiliary vector.  I'm assuming that `unsigned
 120  * long' matches each platform's word length; if this is false then we'll
 121  * need some host-specific tweaking here.
 122  */
 123 union auxval { long i; unsigned long u; const void *p; };
 124 struct auxentry { unsigned long type; union auxval value; };
 125
 126 /* Register each CPU family's interest in the auxiliary vector.  Make sure
 127  * that the necessary entry types are defined.  This is primarily ordered by
 128  * entry type to minimize duplication.
 129  */
 130 #if defined(AT_HWCAP) && CPUFAM_ARMEL
 131 #  define WANT_ANY 1
 132 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 133 #endif
 134
 135 #if defined(AT_HWCAP) && CPUFAM_ARM64
 136 #  define WANT_ANY 1
 137 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 138 #endif
 139
 140 #if defined(AT_HWCAP2) && CPUFAM_ARMEL
 141 #  define WANT_ANY 1
 142 #  define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
 143 #endif
 144
 145 /* If we couldn't find any interesting entries then we can switch all of this
 146  * machinery off.  Also do that if we have no means for atomic updates.
 147  */
 148 #if WANT_ANY && CPU_DISPATCH_P
 149
 150 /* The main output of this section is a bitmask of detected features.  The
 151  * least significant bit will be set if we've tried to probe.  Always access
 152  * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
 153  */
 154 static unsigned hwcaps = 0;
 155
 156 /* For each potentially interesting type which turned out not to exist or be
 157  * wanted, define a dummy macro for the sake of the next step.
 158  */
 159 #ifndef WANT_AT_HWCAP
 160 #  define WANT_AT_HWCAP(_)
 161 #endif
 162 #ifndef WANT_AT_HWCAP2
 163 #  define WANT_AT_HWCAP2(_)
 164 #endif
 165
 166 /* For each CPU family, define two lists.
 167  *
 168  *   * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
 169  *     family tried to register interest in above.  Each entry contains the
 170  *     interesting auxiliary vector entry type, the name of the union branch
 171  *     for its value, and the name of the slot in `struct auxprobe' in which
 172  *     to store the value.
 173  *
 174  *   * `CAPMAP' is a list describing the output features which the CPU family
 175  *     intends to satisfy from the auxiliary vector.  Each entry contains a
 176  *     feature name suffix, and the token name (for `check_env').
 177  */
 178 #if CPUFAM_ARMEL
 179 #  define WANTAUX(_)                                                    \
 180         WANT_AT_HWCAP(_)                                                \
 181         WANT_AT_HWCAP2(_)
 182 #  define CAPMAP(_)                                                     \
 183         _(ARM_VFP, "arm:vfp")                                           \
 184         _(ARM_NEON, "arm:neon")                                         \
 185         _(ARM_V4, "arm:v4")                                             \
 186         _(ARM_D32, "arm:d32")                                           \
 187         _(ARM_AES, "arm:aes")                                           \
 188         _(ARM_PMULL, "arm:pmull")
 189 #endif
 190 #if CPUFAM_ARM64
 191 #  define WANTAUX(_)                                                    \
 192         WANT_AT_HWCAP(_)
 193 #  define CAPMAP(_)                                                     \
 194         _(ARM_AES, "arm:aes")                                           \
 195         _(ARM_PMULL, "arm:pmull")
 196 #endif
 197
 198 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
 199 enum {
 200   HFI_PROBED = 0,
 201 #define HFI__ENUM(feat, tok) HFI_##feat,
 202   CAPMAP(HFI__ENUM)
 203 #undef HFI__ENUM
 204   HFI__END
 205 };
 206 enum {
 207   HF_PROBED = 1,
 208 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
 209   CAPMAP(HF__FLAG)
 210 #undef HF__FLAG
 211   HF__END
 212 };
 213
 214 /* Build a structure in which we can capture the interesting data from the
 215  * auxiliary vector.
 216  */
 217 #define AUXUTYPE_i long
 218 #define AUXUTYPE_u unsigned long
 219 #define AUXUTYPE_p const void *
 220 struct auxprobe {
 221 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
 222   WANTAUX(AUXPROBE__SLOT)
 223 #undef AUXPROBE_SLOT
 224 };
 225
 226 /* --- @probe_hwcaps@ --- *
 227  *
 228  * Arguments:   ---
 229  *
 230  * Returns:     ---
 231  *
 232  * Use:         Attempt to find the auxiliary vector (which is well hidden)
 233  *              and discover interesting features from it.
 234  */
 235
 236 static void probe_hwcaps(void)
 237 {
 238   unsigned hw = HF_PROBED;
 239   struct auxprobe probed = { 0 };
 240
 241   /* Populate `probed' with the information we manage to retrieve from the
 242    * auxiliary vector.  Slots we couldn't find are left zero-valued.
 243    */
 244 #if defined(HAVE_GETAUXVAL)
 245   /* Shiny new libc lets us request individual entry types.  This is almost
 246    * too easy.
 247    */
 248 #  define CAP__GET(type, ubranch, slot)                                 \
 249         probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
 250   WANTAUX(CAP__GET)
 251 #else
 252   /* Otherwise we're a bit stuck, really.  Modern Linux kernels make a copy
 253    * of the vector available in `/procc' so we could try that.
 254    *
 255    * The usual place is stuck on the end of the environment vector, but that
 256    * may well have moved, and we have no way of telling whether it has or
 257    * whether there was ever an auxiliary vector there at all; so don't do
 258    * that.
 259    */
 260   {
 261     FILE *fp = 0;
 262     unsigned char *p = 0, *q = 0;
 263     const struct auxentry *a;
 264     size_t sz, off, n;
 265
 266     /* Open the file and read it into a memory chunk. */
 267     if ((fp = fopen("/proc/self/auxv", "rb")) == 0) goto clean;
 268     sz = 4096; off = 0;
 269     if ((p = malloc(sz)) == 0) goto clean;
 270     for (;;) {
 271       n = fread(p + off, 1, sz - off, fp);
 272       off += n;
 273       if (off < sz) break;
 274       sz *= 2; if ((q = realloc(p, sz)) == 0) break;
 275       p = q;
 276     }
 277
 278     /* Work through the vector (or as much of it as we found) and extract the
 279      * types we're interested in.
 280      */
 281     for (a = (const struct auxentry *)p,
 282            n = sz/sizeof(struct auxentry);
 283          n--; a++) {
 284       switch (a->type) {
 285 #define CAP__SWITCH(type, ubranch, slot)                                \
 286         case type: probed.slot = a->value.ubranch; break;
 287         WANTAUX(CAP__SWITCH)
 288         case AT_NULL: goto clean;
 289       }
 290     }
 291
 292   clean:
 293     if (p) free(p);
 294     if (fp) fclose(fp);
 295   }
 296 #endif
 297
 298   /* Each CPU family now has to pick through what was found and stashed in
 299    * `probed', and set the appropriate flag bits in `hw'.
 300    */
 301 #if CPUFAM_ARMEL
 302   if (probed.hwcap & HWCAP_VFPv3) hw |= HF_ARM_VFP;
 303   if (probed.hwcap & HWCAP_NEON) hw |= HF_ARM_NEON;
 304   if (probed.hwcap & HWCAP_VFPD32) hw |= HF_ARM_D32;
 305   if (probed.hwcap & HWCAP_VFPv4) hw |= HF_ARM_V4;
 306 #  ifdef HWCAP2_AES
 307   if (probed.hwcap2 & HWCAP2_AES) hw |= HF_ARM_AES;
 308 #  endif
 309 #  ifdef HWCAP2_PMULL
 310   if (probed.hwcap2 & HWCAP2_PMULL) hw |= HF_ARM_PMULL;
 311 #  endif
 312 #endif
 313 #if CPUFAM_ARM64
 314   if (probed.hwcap & HWCAP_AES) hw |= HF_ARM_AES;
 315   if (probed.hwcap & HWCAP_PMULL) hw |= HF_ARM_PMULL;
 316 #endif
 317
 318   /* Store the bitmask of features we probed for everyone to see. */
 319   DISPATCH_STORE(hwcaps, hw);
 320
 321   /* Finally, make a report about the things we found.  (Doing this earlier
 322    * will pointlessly widen the window in which multiple threads will do the
 323    * above auxiliary-vector probing.)
 324    */
 325 #define CAP__DEBUG(feat, tok)                                           \
 326   dispatch_debug("check auxv for feature `%s': %s", tok,                \
 327                  hw & HF_##feat ? "available" : "absent");
 328   CAPMAP(CAP__DEBUG)
 329 #undef CAP__DEBUG
 330 }
 331
 332 /* --- @get_hwcaps@ --- *
 333  *
 334  * Arguments:   ---
 335  *
 336  * Returns:     A mask of hardware capabilities and other features, as probed
 337  *              from the auxiliary vector.
 338  */
 339
 340 static unsigned get_hwcaps(void)
 341 {
 342   unsigned hw;
 343
 344   DISPATCH_LOAD(hwcaps, hw);
 345   if (!(hwcaps & HF_PROBED)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps, hw); }
 346   return (hw);
 347 }
 348
 349 #endif
 350
 351 /*----- External interface ------------------------------------------------*/
 352
 353 /* --- @dispatch_debug@ --- *
 354  *
 355  * Arguments:   @const char *fmt@ = a format string
 356  *              @...@ = additional arguments
 357  *
 358  * Returns:     ---
 359  *
 360  * Use:         Writes a formatted message to standard output if dispatch
 361  *              debugging is enabled.
 362  */
 363
 364 void dispatch_debug(const char *fmt, ...)
 365 {
 366   va_list ap;
 367   const char *e = getenv("CATACOMB_CPUDISPATCH_DEBUG");
 368
 369   if (e && *e != 'n' && *e != '0') {
 370     va_start(ap, fmt);
 371     fputs("Catacomb CPUDISPATCH: ", stderr);
 372     vfprintf(stderr, fmt, ap);
 373     fputc('\n', stderr);
 374     va_end(ap);
 375   }
 376 }
 377
 378 /* --- @check_env@ --- *
 379  *
 380  * Arguments:   @const char *ftok@ = feature token
 381  *
 382  * Returns:     Zero if the feature is forced off; positive if it's forced
 383  *              on; negative if the user hasn't decided.
 384  *
 385  * Use:         Checks the environment variable `CATACOMB_CPUFEAT' for the
 386  *              feature token @ftok@.  The variable, if it exists, should be
 387  *              a space-separated sequence of `+tok' and `-tok' items.  These
 388  *              tokens may end in `*', which matches any suffix.
 389  */
 390
 391 static int IGNORABLE check_env(const char *ftok)
 392 {
 393   const char *p, *q, *pp;
 394   int d;
 395
 396   p = getenv("CATACOMB_CPUFEAT");
 397   if (!p) return (-1);
 398
 399   for (;;) {
 400     while (isspace((unsigned char)*p)) p++;
 401     if (!*p) return (-1);
 402     switch (*p) {
 403       case '+': d = +1; p++; break;
 404       case '-': d =  0; p++; break;
 405       default:  d = -1;      break;
 406     }
 407     for (q = p; *q && !isspace((unsigned char)*q); q++);
 408     if (d >= 0) {
 409       for (pp = ftok; p < q && *pp && *p == *pp; p++, pp++);
 410       if ((p == q && !*pp) || (*p == '*' && p + 1 == q)) return (d);
 411     }
 412     p = q;
 413   }
 414   return (-1);
 415 }
 416
 417 /* --- @cpu_feature_p@ --- *
 418  *
 419  * Arguments:   @unsigned feat@ = a @CPUFEAT_...@ code
 420  *
 421  * Returns:     Nonzero if the feature is available.
 422  */
 423
 424 #include <stdio.h>
 425
 426 static int IGNORABLE
 427   feat_debug(const char *ftok, const char *check, int verdict)
 428 {
 429   if (verdict >= 0) {
 430     dispatch_debug("feature `%s': %s -> %s", ftok, check,
 431                    verdict ? "available" : "absent");
 432   }
 433   return (verdict);
 434 }
 435
 436 int cpu_feature_p(int feat)
 437 {
 438   int IGNORABLE f;
 439   IGNORE(f);
 440 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat:             \
 441   if ((f = feat_debug(ftok, "environment override",                     \
 442                       check_env(ftok))) >= 0)                           \
 443     return (f);                                                         \
 444   else                                                                  \
 445     return (feat_debug(ftok, "runtime probe", cond));
 446
 447   switch (feat) {
 448 #if CPUFAM_X86 || CPUFAM_AMD64
 449     CASE_CPUFEAT(X86_SSE2, "x86:sse2",
 450                  cpuid_features_p(CPUID1D_SSE2, 0) &&
 451                  xmm_registers_available_p());
 452     CASE_CPUFEAT(X86_AESNI, "x86:aesni",
 453                  cpuid_features_p(CPUID1D_SSE2, CPUID1C_AESNI) &&
 454                  xmm_registers_available_p());
 455     CASE_CPUFEAT(X86_RDRAND, "x86:rdrand",
 456                  cpuid_features_p(0, CPUID1C_RDRAND));
 457     CASE_CPUFEAT(X86_AVX, "x86:avx",
 458                  cpuid_features_p(0, CPUID1C_AVX) &&
 459                  xmm_registers_available_p());
 460     CASE_CPUFEAT(X86_SSSE3, "x86:ssse3",
 461                  cpuid_features_p(0, CPUID1C_SSSE3) &&
 462                  xmm_registers_available_p());
 463     CASE_CPUFEAT(X86_PCLMUL, "x86:pclmul",
 464                  cpuid_features_p(0, CPUID1C_PCLMUL) &&
 465                  xmm_registers_available_p());
 466 #endif
 467 #ifdef CAPMAP
 468 #  define FEATP__CASE(feat, tok)                                        \
 469         CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
 470     CAPMAP(FEATP__CASE)
 471 #undef FEATP__CASE
 472 #endif
 473     default:
 474       dispatch_debug("denying unknown feature %d", feat);
 475       return (0);
 476   }
 477 #undef CASE_CPUFEAT
 478 }
 479
 480 /*----- That's all, folks -------------------------------------------------*/