Merge branch '2.4.x' into 2.5.x
[catacomb] / 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_PCLMUL (1u << 1)
50 # define CPUID1C_SSSE3 (1u << 9)
51 # define CPUID1C_AESNI (1u << 25)
52 # define CPUID1C_AVX (1u << 28)
53 # define CPUID1C_RDRAND (1u << 30)
54
55 struct cpuid { unsigned a, b, c, d; };
56
57 /* --- @cpuid@ --- *
58 *
59 * Arguments: @struct cpuid *cc@ = where to write the result
60 * @unsigned a, c@ = EAX and ECX registers to set
61 *
62 * Returns: ---
63 *
64 * Use: Minimal C wrapper around the x86 `CPUID' instruction. Checks
65 * that the instruction is actually available before invoking
66 * it; fills the output structure with zero if it's not going to
67 * work.
68 */
69
70 #ifdef __GNUC__
71 # if CPUFAM_X86
72 static __inline__ unsigned getflags(void)
73 { unsigned f; __asm__ ("pushf; popl %0" : "=g" (f)); return (f); }
74 static __inline__ unsigned setflags(unsigned f)
75 {
76 unsigned ff;
77 __asm__ ("pushf; pushl %1; popf; pushf; popl %0; popf"
78 : "=r" (ff)
79 : "r" (f));
80 return (ff);
81 }
82 # else
83 static __inline__ unsigned long getflags(void)
84 { unsigned long f; __asm__ ("pushf; popq %0" : "=g" (f)); return (f); }
85 static __inline__ unsigned long long setflags(unsigned long f)
86 {
87 unsigned long ff;
88 __asm__ ("pushf; pushq %1; popf; pushf; popq %0; popf"
89 : "=r" (ff)
90 : "r" (f));
91 return (ff);
92 }
93 # endif
94 #endif
95
96 static void cpuid(struct cpuid *cc, unsigned a, unsigned c)
97 {
98 #ifdef __GNUC__
99 unsigned f;
100 #endif
101
102 cc->a = cc->b = cc->c = cc->d = 0;
103
104 #ifdef __GNUC__
105 /* Stupid dance to detect whether the CPUID instruction is available. */
106 f = getflags();
107 if (!(setflags(f | EFLAGS_ID) & EFLAGS_ID) ||
108 setflags(f & ~EFLAGS_ID) & EFLAGS_ID) {
109 dispatch_debug("CPUID instruction not available");
110 return;
111 }
112 setflags(f);
113
114 /* Alas, EBX is magical in PIC code, so abuse ESI instead. This isn't
115 * pretty, but it works.
116 */
117 # if CPUFAM_X86
118 __asm__ ("pushl %%ebx; cpuid; movl %%ebx, %%esi; popl %%ebx"
119 : "=a" (cc->a), "=S" (cc->b), "=c" (cc->c), "=d" (cc->d)
120 : "a" (a) , "c" (c));
121 # elif CPUFAM_AMD64
122 __asm__ ("pushq %%rbx; cpuid; movl %%ebx, %%esi; popq %%rbx"
123 : "=a" (cc->a), "=S" (cc->b), "=c" (cc->c), "=d" (cc->d)
124 : "a" (a) , "c" (c));
125 # else
126 # error "I'm confused."
127 # endif
128 dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x",
129 a, c, cc->a, cc->b, cc->c, cc->d);
130 #else
131 dispatch_debug("GNU inline assembler not available; can't CPUID");
132 #endif
133 }
134
135 static unsigned cpuid_maxleaf(void)
136 { struct cpuid c; cpuid(&c, 0, 0); return (c.a); }
137
138 /* --- @cpuid_features_p@ --- *
139 *
140 * Arguments: @unsigned dbits@ = bits to check in EDX
141 * @unsigned cbits@ = bits to check in ECX
142 *
143 * Returns: Nonzero if all the requested bits are set in the CPUID result
144 * on leaf 1.
145 */
146
147 static int cpuid_features_p(unsigned dbits, unsigned cbits)
148 {
149 struct cpuid c;
150 if (cpuid_maxleaf() < 1) return (0);
151 cpuid(&c, 1, 0);
152 return ((c.d & dbits) == dbits && (c.c & cbits) == cbits);
153 }
154
155 /* --- @xmm_registers_available_p@ --- *
156 *
157 * Arguments: ---
158 *
159 * Returns: Nonzero if the operating system has made the XMM registers
160 * available for use.
161 */
162
163 static int xmm_registers_available_p(void)
164 {
165 #ifdef __GNUC__
166 unsigned f;
167 /* This hack is by Agner Fog. Use FXSAVE/FXRSTOR to figure out whether the
168 * XMM registers are actually alive.
169 */
170 if (!cpuid_features_p(CPUID1D_FXSR, 0)) return (0);
171 # if CPUFAM_X86
172 __asm__ ("movl %%esp, %%edx; subl $512, %%esp; andl $~15, %%esp\n"
173 "fxsave (%%esp)\n"
174 "movl 160(%%esp), %%eax; xorl $0xaaaa5555, 160(%%esp)\n"
175 "fxrstor (%%esp); fxsave (%%esp)\n"
176 "movl 160(%%esp), %%ecx; movl %%eax, 160(%%esp)\n"
177 "fxrstor (%%esp); movl %%edx, %%esp\n"
178 "xorl %%ecx, %%eax"
179 : "=a" (f)
180 : /* no inputs */
181 : "%ecx", "%edx");
182 # elif CPUFAM_AMD64
183 __asm__ ("movq %%rsp, %%rdx; subq $512, %%rsp; andq $~15, %%rsp\n"
184 "fxsave (%%rsp)\n"
185 "movl 160(%%rsp), %%eax; xorl $0xaaaa5555, 160(%%rsp)\n"
186 "fxrstor (%%rsp); fxsave (%%rsp)\n"
187 "movl 160(%%rsp), %%ecx; movl %%eax, 160(%%rsp)\n"
188 "fxrstor (%%rsp); movq %%rdx, %%rsp\n"
189 "xorl %%ecx, %%eax"
190 : "=a" (f)
191 : /* no inputs */
192 : "%ecx", "%rdx");
193 # else
194 # error "I'm confused."
195 # endif
196 dispatch_debug("XMM registers %savailable", f ? "" : "not ");
197 return (f);
198 #else
199 dispatch_debug("GNU inline assembler not available; can't check for XMM");
200 return (0);
201 #endif
202 }
203
204 /* --- @rdrand_works_p@ --- *
205 *
206 *
207 * Arguments: ---
208 *
209 * Returns: Nonzero if the `rdrand' instruction actually works. Assumes
210 * that it's already been verified to be safe to issue.
211 */
212
213 #ifdef __GNUC__
214 static int rdrand(unsigned *x)
215 {
216 int i, rc;
217 unsigned _t;
218
219 i = 16;
220 __asm__ ("" : "=g" (_t));
221 __asm__ ("0: rdrand %2; jc 1f; decl %1; jnz 0b\n"
222 "mov $-1, %0; jmp 9f\n"
223 "1: movl %2, (%3); xorl %0, %0\n"
224 "9:"
225 : "=r" (rc), "+r" (i), "+r" (_t)
226 : "r" (x)
227 : "cc");
228 return (rc);
229 }
230 #endif
231
232 static int rdrand_works_p(void)
233 {
234 unsigned ref, x, i;
235
236 /* Check that it doesn't always give the same answer. Try four times: this
237 * will fail with probability %$2^{-128}$% with a truly random generator,
238 * which seems fair enough.
239 */
240 if (rdrand(&ref)) goto fail;
241 for (i = 0; i < 4; i++) {
242 if (rdrand(&x)) goto fail;
243 if (x != ref) goto not_stuck;
244 }
245 dispatch_debug("RDRAND always returns 0x%08x!", ref);
246 return (0);
247
248 not_stuck:
249 dispatch_debug("RDRAND instruction looks plausible");
250 return (1);
251
252 fail:
253 dispatch_debug("RDRAND instruction fails too often");
254 return (0);
255 }
256
257 #endif
258
259 /*----- General feature probing using auxiliary vectors -------------------*/
260
261 /* Try to find the system's definitions for auxiliary vector entries. */
262 #ifdef HAVE_SYS_AUXV_H
263 # include <sys/auxv.h>
264 #endif
265 #ifdef HAVE_LINUX_AUXVEC_H
266 # include <linux/auxvec.h>
267 #endif
268 #ifdef HAVE_ASM_HWCAP_H
269 # include <asm/hwcap.h>
270 #endif
271
272 /* The type of entries in the auxiliary vector. I'm assuming that `unsigned
273 * long' matches each platform's word length; if this is false then we'll
274 * need some host-specific tweaking here.
275 */
276 union auxval { long i; unsigned long u; const void *p; };
277 struct auxentry { unsigned long type; union auxval value; };
278
279 /* Register each CPU family's interest in the auxiliary vector. Make sure
280 * that the necessary entry types are defined. This is primarily ordered by
281 * entry type to minimize duplication.
282 */
283 #if defined(AT_HWCAP) && CPUFAM_ARMEL
284 # define WANT_ANY 1
285 # define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
286 #endif
287
288 #if defined(AT_HWCAP) && CPUFAM_ARM64
289 # define WANT_ANY 1
290 # define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
291 #endif
292
293 #if defined(AT_HWCAP2) && CPUFAM_ARMEL
294 # define WANT_ANY 1
295 # define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
296 #endif
297
298 /* If we couldn't find any interesting entries then we can switch all of this
299 * machinery off. Also do that if we have no means for atomic updates.
300 */
301 #if WANT_ANY && CPU_DISPATCH_P
302
303 /* The main output of this section is a bitmask of detected features. The
304 * least significant bit will be set if we've tried to probe. Always access
305 * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
306 */
307 static unsigned hwcaps = 0;
308
309 /* For each potentially interesting type which turned out not to exist or be
310 * wanted, define a dummy macro for the sake of the next step.
311 */
312 #ifndef WANT_AT_HWCAP
313 # define WANT_AT_HWCAP(_)
314 #endif
315 #ifndef WANT_AT_HWCAP2
316 # define WANT_AT_HWCAP2(_)
317 #endif
318
319 /* For each CPU family, define two lists.
320 *
321 * * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
322 * family tried to register interest in above. Each entry contains the
323 * interesting auxiliary vector entry type, the name of the union branch
324 * for its value, and the name of the slot in `struct auxprobe' in which
325 * to store the value.
326 *
327 * * `CAPMAP' is a list describing the output features which the CPU family
328 * intends to satisfy from the auxiliary vector. Each entry contains a
329 * feature name suffix, and the token name (for `check_env').
330 */
331 #if CPUFAM_ARMEL
332 # define WANTAUX(_) \
333 WANT_AT_HWCAP(_) \
334 WANT_AT_HWCAP2(_)
335 # define CAPMAP(_) \
336 _(ARM_VFP, "arm:vfp") \
337 _(ARM_NEON, "arm:neon") \
338 _(ARM_V4, "arm:v4") \
339 _(ARM_D32, "arm:d32") \
340 _(ARM_AES, "arm:aes") \
341 _(ARM_PMULL, "arm:pmull")
342 #endif
343 #if CPUFAM_ARM64
344 # define WANTAUX(_) \
345 WANT_AT_HWCAP(_)
346 # define CAPMAP(_) \
347 _(ARM_AES, "arm:aes") \
348 _(ARM_PMULL, "arm:pmull")
349 #endif
350
351 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
352 enum {
353 HFI_PROBED = 0,
354 #define HFI__ENUM(feat, tok) HFI_##feat,
355 CAPMAP(HFI__ENUM)
356 #undef HFI__ENUM
357 HFI__END
358 };
359 enum {
360 HF_PROBED = 1,
361 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
362 CAPMAP(HF__FLAG)
363 #undef HF__FLAG
364 HF__END
365 };
366
367 /* Build a structure in which we can capture the interesting data from the
368 * auxiliary vector.
369 */
370 #define AUXUTYPE_i long
371 #define AUXUTYPE_u unsigned long
372 #define AUXUTYPE_p const void *
373 struct auxprobe {
374 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
375 WANTAUX(AUXPROBE__SLOT)
376 #undef AUXPROBE_SLOT
377 };
378
379 /* --- @probe_hwcaps@ --- *
380 *
381 * Arguments: ---
382 *
383 * Returns: ---
384 *
385 * Use: Attempt to find the auxiliary vector (which is well hidden)
386 * and discover interesting features from it.
387 */
388
389 static void probe_hwcaps(void)
390 {
391 unsigned hw = HF_PROBED;
392 struct auxprobe probed = { 0 };
393
394 /* Populate `probed' with the information we manage to retrieve from the
395 * auxiliary vector. Slots we couldn't find are left zero-valued.
396 */
397 #if defined(HAVE_GETAUXVAL)
398 /* Shiny new libc lets us request individual entry types. This is almost
399 * too easy.
400 */
401 # define CAP__GET(type, ubranch, slot) \
402 probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
403 WANTAUX(CAP__GET)
404 #else
405 /* Otherwise we're a bit stuck, really. Modern Linux kernels make a copy
406 * of the vector available in `/procc' so we could try that.
407 *
408 * The usual place is stuck on the end of the environment vector, but that
409 * may well have moved, and we have no way of telling whether it has or
410 * whether there was ever an auxiliary vector there at all; so don't do
411 * that.
412 */
413 {
414 FILE *fp = 0;
415 unsigned char *p = 0, *q = 0;
416 const struct auxentry *a;
417 size_t sz, off, n;
418
419 /* Open the file and read it into a memory chunk. */
420 if ((fp = fopen("/proc/self/auxv", "rb")) == 0) goto clean;
421 sz = 4096; off = 0;
422 if ((p = malloc(sz)) == 0) goto clean;
423 for (;;) {
424 n = fread(p + off, 1, sz - off, fp);
425 off += n;
426 if (off < sz) break;
427 sz *= 2; if ((q = realloc(p, sz)) == 0) break;
428 p = q;
429 }
430
431 /* Work through the vector (or as much of it as we found) and extract the
432 * types we're interested in.
433 */
434 for (a = (const struct auxentry *)p,
435 n = sz/sizeof(struct auxentry);
436 n--; a++) {
437 switch (a->type) {
438 #define CAP__SWITCH(type, ubranch, slot) \
439 case type: probed.slot = a->value.ubranch; break;
440 WANTAUX(CAP__SWITCH)
441 case AT_NULL: goto clean;
442 }
443 }
444
445 clean:
446 if (p) free(p);
447 if (fp) fclose(fp);
448 }
449 #endif
450
451 /* Each CPU family now has to pick through what was found and stashed in
452 * `probed', and set the appropriate flag bits in `hw'.
453 */
454 #if CPUFAM_ARMEL
455 if (probed.hwcap & HWCAP_VFPv3) hw |= HF_ARM_VFP;
456 if (probed.hwcap & HWCAP_NEON) hw |= HF_ARM_NEON;
457 if (probed.hwcap & HWCAP_VFPD32) hw |= HF_ARM_D32;
458 if (probed.hwcap & HWCAP_VFPv4) hw |= HF_ARM_V4;
459 # ifdef HWCAP2_AES
460 if (probed.hwcap2 & HWCAP2_AES) hw |= HF_ARM_AES;
461 # endif
462 # ifdef HWCAP2_PMULL
463 if (probed.hwcap2 & HWCAP2_PMULL) hw |= HF_ARM_PMULL;
464 # endif
465 #endif
466 #if CPUFAM_ARM64
467 if (probed.hwcap & HWCAP_AES) hw |= HF_ARM_AES;
468 if (probed.hwcap & HWCAP_PMULL) hw |= HF_ARM_PMULL;
469 #endif
470
471 /* Store the bitmask of features we probed for everyone to see. */
472 DISPATCH_STORE(hwcaps, hw);
473
474 /* Finally, make a report about the things we found. (Doing this earlier
475 * will pointlessly widen the window in which multiple threads will do the
476 * above auxiliary-vector probing.)
477 */
478 #define CAP__DEBUG(feat, tok) \
479 dispatch_debug("check auxv for feature `%s': %s", tok, \
480 hw & HF_##feat ? "available" : "absent");
481 CAPMAP(CAP__DEBUG)
482 #undef CAP__DEBUG
483 }
484
485 /* --- @get_hwcaps@ --- *
486 *
487 * Arguments: ---
488 *
489 * Returns: A mask of hardware capabilities and other features, as probed
490 * from the auxiliary vector.
491 */
492
493 static unsigned get_hwcaps(void)
494 {
495 unsigned hw;
496
497 DISPATCH_LOAD(hwcaps, hw);
498 if (!(hwcaps & HF_PROBED)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps, hw); }
499 return (hw);
500 }
501
502 #endif
503
504 /*----- External interface ------------------------------------------------*/
505
506 /* --- @dispatch_debug@ --- *
507 *
508 * Arguments: @const char *fmt@ = a format string
509 * @...@ = additional arguments
510 *
511 * Returns: ---
512 *
513 * Use: Writes a formatted message to standard output if dispatch
514 * debugging is enabled.
515 */
516
517 void dispatch_debug(const char *fmt, ...)
518 {
519 va_list ap;
520 const char *e = getenv("CATACOMB_CPUDISPATCH_DEBUG");
521
522 if (e && *e != 'n' && *e != '0') {
523 va_start(ap, fmt);
524 fputs("Catacomb CPUDISPATCH: ", stderr);
525 vfprintf(stderr, fmt, ap);
526 fputc('\n', stderr);
527 va_end(ap);
528 }
529 }
530
531 /* --- @check_env@ --- *
532 *
533 * Arguments: @const char *ftok@ = feature token
534 *
535 * Returns: Zero if the feature is forced off; positive if it's forced
536 * on; negative if the user hasn't decided.
537 *
538 * Use: Checks the environment variable `CATACOMB_CPUFEAT' for the
539 * feature token @ftok@. The variable, if it exists, should be
540 * a space-separated sequence of `+tok' and `-tok' items. These
541 * tokens may end in `*', which matches any suffix.
542 */
543
544 static int IGNORABLE check_env(const char *ftok)
545 {
546 const char *p, *q, *pp;
547 int d;
548
549 p = getenv("CATACOMB_CPUFEAT");
550 if (!p) return (-1);
551
552 for (;;) {
553 while (isspace((unsigned char)*p)) p++;
554 if (!*p) return (-1);
555 switch (*p) {
556 case '+': d = +1; p++; break;
557 case '-': d = 0; p++; break;
558 default: d = -1; break;
559 }
560 for (q = p; *q && !isspace((unsigned char)*q); q++);
561 if (d >= 0) {
562 for (pp = ftok; p < q && *pp && *p == *pp; p++, pp++);
563 if ((p == q && !*pp) || (*p == '*' && p + 1 == q)) return (d);
564 }
565 p = q;
566 }
567 return (-1);
568 }
569
570 /* --- @cpu_feature_p@ --- *
571 *
572 * Arguments: @unsigned feat@ = a @CPUFEAT_...@ code
573 *
574 * Returns: Nonzero if the feature is available.
575 */
576
577 #include <stdio.h>
578
579 static int IGNORABLE
580 feat_debug(const char *ftok, const char *check, int verdict)
581 {
582 if (verdict >= 0) {
583 dispatch_debug("feature `%s': %s -> %s", ftok, check,
584 verdict ? "available" : "absent");
585 }
586 return (verdict);
587 }
588
589 int cpu_feature_p(int feat)
590 {
591 int IGNORABLE f;
592 IGNORE(f);
593 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat: \
594 if ((f = feat_debug(ftok, "environment override", \
595 check_env(ftok))) >= 0) \
596 return (f); \
597 else \
598 return (feat_debug(ftok, "runtime probe", cond));
599
600 switch (feat) {
601 #if CPUFAM_X86 || CPUFAM_AMD64
602 CASE_CPUFEAT(X86_SSE2, "x86:sse2",
603 cpuid_features_p(CPUID1D_SSE2, 0) &&
604 xmm_registers_available_p());
605 CASE_CPUFEAT(X86_AESNI, "x86:aesni",
606 cpuid_features_p(CPUID1D_SSE2, CPUID1C_AESNI) &&
607 xmm_registers_available_p());
608 CASE_CPUFEAT(X86_RDRAND, "x86:rdrand",
609 cpuid_features_p(0, CPUID1C_RDRAND) && rdrand_works_p());
610 CASE_CPUFEAT(X86_AVX, "x86:avx",
611 cpuid_features_p(0, CPUID1C_AVX) &&
612 xmm_registers_available_p());
613 CASE_CPUFEAT(X86_SSSE3, "x86:ssse3",
614 cpuid_features_p(0, CPUID1C_SSSE3) &&
615 xmm_registers_available_p());
616 CASE_CPUFEAT(X86_PCLMUL, "x86:pclmul",
617 cpuid_features_p(0, CPUID1C_PCLMUL) &&
618 xmm_registers_available_p());
619 #endif
620 #ifdef CAPMAP
621 # define FEATP__CASE(feat, tok) \
622 CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
623 CAPMAP(FEATP__CASE)
624 #undef FEATP__CASE
625 #endif
626 default:
627 dispatch_debug("denying unknown feature %d", feat);
628 return (0);
629 }
630 #undef CASE_CPUFEAT
631 }
632
633 /*----- That's all, folks -------------------------------------------------*/