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 #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_HWCAP2) && CPUFAM_ARMEL
 233 #  define WANT_ANY 1
 234 #  define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
 235 #endif
 236
 237 /* If we couldn't find any interesting entries then we can switch all of this
 238  * machinery off.  Also do that if we have no means for atomic updates.
 239  */
 240 #if WANT_ANY && CPU_DISPATCH_P
 241
 242 /* The main output of this section is a bitmask of detected features.  The
 243  * least significant bit will be set if we've tried to probe.  Always access
 244  * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
 245  */
 246 static unsigned hwcaps = 0;
 247
 248 /* For each potentially interesting type which turned out not to exist or be
 249  * wanted, define a dummy macro for the sake of the next step.
 250  */
 251 #ifndef WANT_AT_HWCAP
 252 #  define WANT_AT_HWCAP(_)
 253 #endif
 254 #ifndef WANT_AT_HWCAP2
 255 #  define WANT_AT_HWCAP2(_)
 256 #endif
 257
 258 /* For each CPU family, define two lists.
 259  *
 260  *   * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
 261  *     family tried to register interest in above.  Each entry contains the
 262  *     interesting auxiliary vector entry type, the name of the union branch
 263  *     for its value, and the name of the slot in `struct auxprobe' in which
 264  *     to store the value.
 265  *
 266  *   * `CAPMAP' is a list describing the output features which the CPU family
 267  *     intends to satisfy from the auxiliary vector.  Each entry contains a
 268  *     feature name suffix, and the token name (for `check_env').
 269  */
 270 #if CPUFAM_ARMEL
 271 #  define WANTAUX(_)                                                    \
 272         WANT_AT_HWCAP(_)                                                \
 273         WANT_AT_HWCAP2(_)
 274 #  define CAPMAP(_)                                                     \
 275         _(ARM_VFP, "arm:vfp")                                           \
 276         _(ARM_NEON, "arm:neon")                                         \
 277         _(ARM_V4, "arm:v4")                                             \
 278         _(ARM_D32, "arm:d32")                                           \
 279         _(ARM_AES, "arm:aes")
 280 #endif
 281
 282 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
 283 enum {
 284   HFI_PROBED = 0,
 285 #define HFI__ENUM(feat, tok) HFI_##feat,
 286   CAPMAP(HFI__ENUM)
 287 #undef HFI__ENUM
 288   HFI__END
 289 };
 290 enum {
 291   HF_PROBED = 1,
 292 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
 293   CAPMAP(HF__FLAG)
 294 #undef HF__FLAG
 295   HF__END
 296 };
 297
 298 /* Build a structure in which we can capture the interesting data from the
 299  * auxiliary vector.
 300  */
 301 #define AUXUTYPE_i long
 302 #define AUXUTYPE_u unsigned long
 303 #define AUXUTYPE_p const void *
 304 struct auxprobe {
 305 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
 306   WANTAUX(AUXPROBE__SLOT)
 307 #undef AUXPROBE_SLOT
 308 };
 309
 310 /* --- @probe_hwcaps@ --- *
 311  *
 312  * Arguments:   ---
 313  *
 314  * Returns:     ---
 315  *
 316  * Use:         Attempt to find the auxiliary vector (which is well hidden)
 317  *              and discover interesting features from it.
 318  */
 319
 320 static void probe_hwcaps(void)
 321 {
 322   unsigned hw = HF_PROBED;
 323   struct auxprobe probed = { 0 };
 324
 325   /* Populate `probed' with the information we manage to retrieve from the
 326    * auxiliary vector.  Slots we couldn't find are left zero-valued.
 327    */
 328 #if defined(HAVE_GETAUXVAL)
 329   /* Shiny new libc lets us request individual entry types.  This is almost
 330    * too easy.
 331    */
 332 #  define CAP__GET(type, ubranch, slot)                                 \
 333         probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
 334   WANTAUX(CAP__GET)
 335 #else
 336   /* Otherwise we're a bit stuck, really.  Modern Linux kernels make a copy
 337    * of the vector available in `/procc' so we could try that.
 338    *
 339    * The usual place is stuck on the end of the environment vector, but that
 340    * may well have moved, and we have no way of telling whether it has or
 341    * whether there was ever an auxiliary vector there at all; so don't do
 342    * that.
 343    */
 344   {
 345     FILE *fp = 0;
 346     unsigned char *p = 0, *q = 0;
 347     const struct auxentry *a;
 348     size_t sz, off, n;
 349
 350     /* Open the file and read it into a memory chunk. */
 351     if ((fp = fopen("/proc/self/auxv", "rb")) == 0) goto clean;
 352     sz = 4096; off = 0;
 353     if ((p = malloc(sz)) == 0) goto clean;
 354     for (;;) {
 355       n = fread(p + off, 1, sz - off, fp);
 356       off += n;
 357       if (off < sz) break;
 358       sz *= 2; if ((q = realloc(p, sz)) == 0) break;
 359       p = q;
 360     }
 361
 362     /* Work through the vector (or as much of it as we found) and extract the
 363      * types we're interested in.
 364      */
 365     for (a = (const struct auxentry *)p,
 366            n = sz/sizeof(struct auxentry);
 367          n--; a++) {
 368       switch (a->type) {
 369 #define CAP__SWITCH(type, ubranch, slot)                                \
 370         case type: probed.slot = a->value.ubranch; break;
 371         WANTAUX(CAP__SWITCH)
 372         case AT_NULL: goto clean;
 373       }
 374     }
 375
 376   clean:
 377     if (p) free(p);
 378     if (fp) fclose(fp);
 379   }
 380 #endif
 381
 382   /* Each CPU family now has to pick through what was found and stashed in
 383    * `probed', and set the appropriate flag bits in `hw'.
 384    */
 385 #if CPUFAM_ARMEL
 386   if (probed.hwcap & HWCAP_VFPv3) hw |= HF_ARM_VFP;
 387   if (probed.hwcap & HWCAP_NEON) hw |= HF_ARM_NEON;
 388   if (probed.hwcap & HWCAP_VFPD32) hw |= HF_ARM_D32;
 389   if (probed.hwcap & HWCAP_VFPv4) hw |= HF_ARM_V4;
 390 #  ifdef HWCAP2_AES
 391   if (probed.hwcap2 & HWCAP2_AES) hw |= HF_ARM_AES;
 392 #  endif
 393 #endif
 394
 395   /* Store the bitmask of features we probed for everyone to see. */
 396   DISPATCH_STORE(hwcaps, hw);
 397
 398   /* Finally, make a report about the things we found.  (Doing this earlier
 399    * will pointlessly widen the window in which multiple threads will do the
 400    * above auxiliary-vector probing.)
 401    */
 402 #define CAP__DEBUG(feat, tok)                                           \
 403   dispatch_debug("check auxv for feature `%s': %s", tok,                \
 404                  hw & HF_##feat ? "available" : "absent");
 405   CAPMAP(CAP__DEBUG)
 406 #undef CAP__DEBUG
 407 }
 408
 409 /* --- @get_hwcaps@ --- *
 410  *
 411  * Arguments:   ---
 412  *
 413  * Returns:     A mask of hardware capabilities and other features, as probed
 414  *              from the auxiliary vector.
 415  */
 416
 417 static unsigned get_hwcaps(void)
 418 {
 419   unsigned hw;
 420
 421   DISPATCH_LOAD(hwcaps, hw);
 422   if (!(hwcaps & HF_PROBED)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps, hw); }
 423   return (hw);
 424 }
 425
 426 #endif
 427
 428 /*----- External interface ------------------------------------------------*/
 429
 430 /* --- @dispatch_debug@ --- *
 431  *
 432  * Arguments:   @const char *fmt@ = a format string
 433  *              @...@ = additional arguments
 434  *
 435  * Returns:     ---
 436  *
 437  * Use:         Writes a formatted message to standard output if dispatch
 438  *              debugging is enabled.
 439  */
 440
 441 void dispatch_debug(const char *fmt, ...)
 442 {
 443   va_list ap;
 444   const char *e = getenv("CATACOMB_CPUDISPATCH_DEBUG");
 445
 446   if (e && *e != 'n' && *e != '0') {
 447     va_start(ap, fmt);
 448     fputs("Catacomb CPUDISPATCH: ", stderr);
 449     vfprintf(stderr, fmt, ap);
 450     fputc('\n', stderr);
 451     va_end(ap);
 452   }
 453 }
 454
 455 /* --- @check_env@ --- *
 456  *
 457  * Arguments:   @const char *ftok@ = feature token
 458  *
 459  * Returns:     Zero if the feature is forced off; positive if it's forced
 460  *              on; negative if the user hasn't decided.
 461  *
 462  * Use:         Checks the environment variable `CATACOMB_CPUFEAT' for the
 463  *              feature token @ftok@.  The variable, if it exists, should be
 464  *              a space-separated sequence of `+tok' and `-tok' items.  These
 465  *              tokens may end in `*', which matches any suffix.
 466  */
 467
 468 static int IGNORABLE check_env(const char *ftok)
 469 {
 470   const char *p, *q, *pp;
 471   int d;
 472
 473   p = getenv("CATACOMB_CPUFEAT");
 474   if (!p) return (-1);
 475
 476   for (;;) {
 477     while (isspace((unsigned char)*p)) p++;
 478     if (!*p) return (-1);
 479     switch (*p) {
 480       case '+': d = +1; p++; break;
 481       case '-': d =  0; p++; break;
 482       default:  d = -1;      break;
 483     }
 484     for (q = p; *q && !isspace((unsigned char)*q); q++);
 485     if (d >= 0) {
 486       for (pp = ftok; p < q && *pp && *p == *pp; p++, pp++);
 487       if ((p == q && !*pp) || (*p == '*' && p + 1 == q)) return (d);
 488     }
 489     p = q;
 490   }
 491   return (-1);
 492 }
 493
 494 /* --- @cpu_feature_p@ --- *
 495  *
 496  * Arguments:   @unsigned feat@ = a @CPUFEAT_...@ code
 497  *
 498  * Returns:     Nonzero if the feature is available.
 499  */
 500
 501 #include <stdio.h>
 502
 503 static int IGNORABLE
 504   feat_debug(const char *ftok, const char *check, int verdict)
 505 {
 506   if (verdict >= 0) {
 507     dispatch_debug("feature `%s': %s -> %s", ftok, check,
 508                    verdict ? "available" : "absent");
 509   }
 510   return (verdict);
 511 }
 512
 513 int cpu_feature_p(int feat)
 514 {
 515   int IGNORABLE f;
 516   IGNORE(f);
 517 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat:             \
 518   if ((f = feat_debug(ftok, "environment override",                     \
 519                       check_env(ftok))) >= 0)                           \
 520     return (f);                                                         \
 521   else                                                                  \
 522     return (feat_debug(ftok, "runtime probe", cond));
 523
 524   switch (feat) {
 525 #if CPUFAM_X86 || CPUFAM_AMD64
 526     CASE_CPUFEAT(X86_SSE2, "x86:sse2",
 527                  xmm_registers_available_p() &&
 528                  cpuid_features_p(CPUID1D_SSE2, 0));
 529     CASE_CPUFEAT(X86_AESNI, "x86:aesni",
 530                  xmm_registers_available_p() &&
 531                  cpuid_features_p(CPUID1D_SSE2, CPUID1C_AESNI));
 532     CASE_CPUFEAT(X86_RDRAND, "x86:rdrand",
 533                  cpuid_features_p(0, CPUID1C_RDRAND));
 534 #endif
 535 #ifdef CAPMAP
 536 #  define FEATP__CASE(feat, tok)                                        \
 537         CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
 538     CAPMAP(FEATP__CASE)
 539 #undef FEATP__CASE
 540 #endif
 541     default:
 542       dispatch_debug("denying unknown feature %d", feat);
 543       return (0);
 544   }
 545 #undef CASE_CPUFEAT
 546 }
 547
 548 /*----- That's all, folks -------------------------------------------------*/