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