9ba6a7cd3cd79b55b298ed6fb9f39d2de1de5128
3 * CPU-specific dispatch
5 * (c) 2015 Straylight/Edgeware
8 /*----- Licensing notice --------------------------------------------------*
10 * This file is part of Catacomb.
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.
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.
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,
28 /*----- Header files ------------------------------------------------------*/
38 #include <mLib/macros.h>
42 /*----- Intel x86/AMD64 feature probing -----------------------------------*/
44 #if CPUFAM_X86 || CPUFAM_AMD64
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_AVX (1u << 28)
51 # define CPUID1C_RDRAND (1u << 30)
53 struct cpuid
{ unsigned a
, b
, c
, d
; };
57 * Arguments: @struct cpuid *cc@ = where to write the result
58 * @unsigned a, c@ = EAX and ECX registers to set
62 * Use: Minimal C wrapper around the x86 `CPUID' instruction. Checks
63 * that the instruction is actually available before invoking
64 * it; fills the output structure with zero if it's not going to
70 static __inline__
unsigned getflags(void)
71 { unsigned f
; __asm__ ("pushf; popl %0" : "=g" (f
)); return (f
); }
72 static __inline__
unsigned setflags(unsigned f
)
75 __asm__ ("pushf; pushl %1; popf; pushf; popl %0; popf"
81 static __inline__
unsigned long getflags(void)
82 { unsigned long f
; __asm__ ("pushf; popq %0" : "=g" (f
)); return (f
); }
83 static __inline__
unsigned long long setflags(unsigned long f
)
86 __asm__ ("pushf; pushq %1; popf; pushf; popq %0; popf"
94 static void cpuid(struct cpuid
*cc
, unsigned a
, unsigned c
)
100 cc
->a
= cc
->b
= cc
->c
= cc
->d
= 0;
103 /* Stupid dance to detect whether the CPUID instruction is available. */
105 if (!(setflags(f
| EFLAGS_ID
) & EFLAGS_ID
) ||
106 setflags(f
& ~EFLAGS_ID
) & EFLAGS_ID
) {
107 dispatch_debug("CPUID instruction not available");
112 /* Alas, EBX is magical in PIC code, so abuse ESI instead. This isn't
113 * pretty, but it works.
116 __asm__ ("pushl %%ebx; cpuid; movl %%ebx, %%esi; popl %%ebx"
117 : "=a" (cc
->a
), "=S" (cc
->b
), "=c" (cc
->c
), "=d" (cc
->d
)
118 : "a" (a
) , "c" (c
));
120 __asm__ ("pushq %%rbx; cpuid; movl %%ebx, %%esi; popq %%rbx"
121 : "=a" (cc
->a
), "=S" (cc
->b
), "=c" (cc
->c
), "=d" (cc
->d
)
122 : "a" (a
) , "c" (c
));
124 # error "I'm confused."
126 dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x",
127 a
, c
, cc
->a
, cc
->b
, cc
->c
, cc
->d
);
129 dispatch_debug("GNU inline assembler not available; can't CPUID");
133 static unsigned cpuid_maxleaf(void)
134 { struct cpuid c
; cpuid(&c
, 0, 0); return (c
.a
); }
136 /* --- @cpuid_features_p@ --- *
138 * Arguments: @unsigned dbits@ = bits to check in EDX
139 * @unsigned cbits@ = bits to check in ECX
141 * Returns: Nonzero if all the requested bits are set in the CPUID result
145 static int cpuid_features_p(unsigned dbits
, unsigned cbits
)
148 if (cpuid_maxleaf() < 1) return (0);
150 return ((c
.d
& dbits
) == dbits
&& (c
.c
& cbits
) == cbits
);
153 /* --- @xmm_registers_available_p@ --- *
157 * Returns: Nonzero if the operating system has made the XMM registers
161 static int xmm_registers_available_p(void)
165 /* This hack is by Agner Fog. Use FXSAVE/FXRSTOR to figure out whether the
166 * XMM registers are actually alive.
168 if (!cpuid_features_p(CPUID1D_FXSR
, 0)) return (0);
170 __asm__ ("movl %%esp, %%edx; subl $512, %%esp; andl $~15, %%esp\n"
172 "movl 160(%%esp), %%eax; xorl $0xaaaa5555, 160(%%esp)\n"
173 "fxrstor (%%esp); fxsave (%%esp)\n"
174 "movl 160(%%esp), %%ecx; movl %%eax, 160(%%esp)\n"
175 "fxrstor (%%esp); movl %%edx, %%esp\n"
181 __asm__ ("movq %%rsp, %%rdx; subq $512, %%rsp; andq $~15, %%rsp\n"
183 "movl 160(%%rsp), %%eax; xorl $0xaaaa5555, 160(%%rsp)\n"
184 "fxrstor (%%rsp); fxsave (%%rsp)\n"
185 "movl 160(%%rsp), %%ecx; movl %%eax, 160(%%rsp)\n"
186 "fxrstor (%%rsp); movq %%rdx, %%rsp\n"
192 # error "I'm confused."
194 dispatch_debug("XMM registers %savailable", f ?
"" : "not ");
197 dispatch_debug("GNU inline assembler not available; can't check for XMM");
204 /*----- General feature probing using auxiliary vectors -------------------*/
206 /* Try to find the system's definitions for auxiliary vector entries. */
207 #ifdef HAVE_SYS_AUXV_H
208 # include <sys/auxv.h>
210 #ifdef HAVE_LINUX_AUXVEC_H
211 # include <linux/auxvec.h>
213 #ifdef HAVE_ASM_HWCAP_H
214 # include <asm/hwcap.h>
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.
221 union auxval
{ long i
; unsigned long u
; const void *p
; };
222 struct auxentry
{ unsigned long type
; union auxval value
; };
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.
228 #if defined(AT_HWCAP) && CPUFAM_ARMEL
230 # define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
233 #if defined(AT_HWCAP) && CPUFAM_ARM64
235 # define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
238 #if defined(AT_HWCAP2) && CPUFAM_ARMEL
240 # define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
243 /* If we couldn't find any interesting entries then we can switch all of this
244 * machinery off. Also do that if we have no means for atomic updates.
246 #if WANT_ANY && CPU_DISPATCH_P
248 /* The main output of this section is a bitmask of detected features. The
249 * least significant bit will be set if we've tried to probe. Always access
250 * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
252 static unsigned hwcaps
= 0;
254 /* For each potentially interesting type which turned out not to exist or be
255 * wanted, define a dummy macro for the sake of the next step.
257 #ifndef WANT_AT_HWCAP
258 # define WANT_AT_HWCAP(_)
260 #ifndef WANT_AT_HWCAP2
261 # define WANT_AT_HWCAP2(_)
264 /* For each CPU family, define two lists.
266 * * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
267 * family tried to register interest in above. Each entry contains the
268 * interesting auxiliary vector entry type, the name of the union branch
269 * for its value, and the name of the slot in `struct auxprobe' in which
270 * to store the value.
272 * * `CAPMAP' is a list describing the output features which the CPU family
273 * intends to satisfy from the auxiliary vector. Each entry contains a
274 * feature name suffix, and the token name (for `check_env').
277 # define WANTAUX(_) \
281 _(ARM_VFP, "arm:vfp") \
282 _(ARM_NEON, "arm:neon") \
283 _(ARM_V4, "arm:v4") \
284 _(ARM_D32, "arm:d32") \
285 _(ARM_AES, "arm:aes")
288 # define WANTAUX(_) \
291 _(ARM_AES, "arm:aes")
294 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
297 #define HFI__ENUM(feat, tok) HFI_##feat,
304 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
310 /* Build a structure in which we can capture the interesting data from the
313 #define AUXUTYPE_i long
314 #define AUXUTYPE_u unsigned long
315 #define AUXUTYPE_p const void *
317 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
318 WANTAUX(AUXPROBE__SLOT
)
322 /* --- @probe_hwcaps@ --- *
328 * Use: Attempt to find the auxiliary vector (which is well hidden)
329 * and discover interesting features from it.
332 static void probe_hwcaps(void)
334 unsigned hw
= HF_PROBED
;
335 struct auxprobe probed
= { 0 };
337 /* Populate `probed' with the information we manage to retrieve from the
338 * auxiliary vector. Slots we couldn't find are left zero-valued.
340 #if defined(HAVE_GETAUXVAL)
341 /* Shiny new libc lets us request individual entry types. This is almost
344 # define CAP__GET(type, ubranch, slot) \
345 probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
348 /* Otherwise we're a bit stuck, really. Modern Linux kernels make a copy
349 * of the vector available in `/procc' so we could try that.
351 * The usual place is stuck on the end of the environment vector, but that
352 * may well have moved, and we have no way of telling whether it has or
353 * whether there was ever an auxiliary vector there at all; so don't do
358 unsigned char *p
= 0, *q
= 0;
359 const struct auxentry
*a
;
362 /* Open the file and read it into a memory chunk. */
363 if ((fp
= fopen("/proc/self/auxv", "rb")) == 0) goto clean
;
365 if ((p
= malloc(sz
)) == 0) goto clean
;
367 n
= fread(p
+ off
, 1, sz
- off
, fp
);
370 sz
*= 2; if ((q
= realloc(p
, sz
)) == 0) break;
374 /* Work through the vector (or as much of it as we found) and extract the
375 * types we're interested in.
377 for (a
= (const struct auxentry
*)p
,
378 n
= sz
/sizeof(struct auxentry
);
381 #define CAP__SWITCH(type, ubranch, slot) \
382 case type: probed.slot = a->value.ubranch; break;
384 case AT_NULL
: goto clean
;
394 /* Each CPU family now has to pick through what was found and stashed in
395 * `probed', and set the appropriate flag bits in `hw'.
398 if (probed
.hwcap
& HWCAP_VFPv3
) hw
|= HF_ARM_VFP
;
399 if (probed
.hwcap
& HWCAP_NEON
) hw
|= HF_ARM_NEON
;
400 if (probed
.hwcap
& HWCAP_VFPD32
) hw
|= HF_ARM_D32
;
401 if (probed
.hwcap
& HWCAP_VFPv4
) hw
|= HF_ARM_V4
;
403 if (probed
.hwcap2
& HWCAP2_AES
) hw
|= HF_ARM_AES
;
407 if (probed
.hwcap
& HWCAP_AES
) hw
|= HF_ARM_AES
;
410 /* Store the bitmask of features we probed for everyone to see. */
411 DISPATCH_STORE(hwcaps
, hw
);
413 /* Finally, make a report about the things we found. (Doing this earlier
414 * will pointlessly widen the window in which multiple threads will do the
415 * above auxiliary-vector probing.)
417 #define CAP__DEBUG(feat, tok) \
418 dispatch_debug("check auxv for feature `%s': %s", tok, \
419 hw & HF_##feat ? "available" : "absent");
424 /* --- @get_hwcaps@ --- *
428 * Returns: A mask of hardware capabilities and other features, as probed
429 * from the auxiliary vector.
432 static unsigned get_hwcaps(void)
436 DISPATCH_LOAD(hwcaps
, hw
);
437 if (!(hwcaps
& HF_PROBED
)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps
, hw
); }
443 /*----- External interface ------------------------------------------------*/
445 /* --- @dispatch_debug@ --- *
447 * Arguments: @const char *fmt@ = a format string
448 * @...@ = additional arguments
452 * Use: Writes a formatted message to standard output if dispatch
453 * debugging is enabled.
456 void dispatch_debug(const char *fmt
, ...)
459 const char *e
= getenv("CATACOMB_CPUDISPATCH_DEBUG");
461 if (e
&& *e
!= 'n' && *e
!= '0') {
463 fputs("Catacomb CPUDISPATCH: ", stderr
);
464 vfprintf(stderr
, fmt
, ap
);
470 /* --- @check_env@ --- *
472 * Arguments: @const char *ftok@ = feature token
474 * Returns: Zero if the feature is forced off; positive if it's forced
475 * on; negative if the user hasn't decided.
477 * Use: Checks the environment variable `CATACOMB_CPUFEAT' for the
478 * feature token @ftok@. The variable, if it exists, should be
479 * a space-separated sequence of `+tok' and `-tok' items. These
480 * tokens may end in `*', which matches any suffix.
483 static int IGNORABLE
check_env(const char *ftok
)
485 const char *p
, *q
, *pp
;
488 p
= getenv("CATACOMB_CPUFEAT");
492 while (isspace((unsigned char)*p
)) p
++;
493 if (!*p
) return (-1);
495 case '+': d
= +1; p
++; break;
496 case '-': d
= 0; p
++; break;
497 default: d
= -1; break;
499 for (q
= p
; *q
&& !isspace((unsigned char)*q
); q
++);
501 for (pp
= ftok
; p
< q
&& *pp
&& *p
== *pp
; p
++, pp
++);
502 if ((p
== q
&& !*pp
) || (*p
== '*' && p
+ 1 == q
)) return (d
);
509 /* --- @cpu_feature_p@ --- *
511 * Arguments: @unsigned feat@ = a @CPUFEAT_...@ code
513 * Returns: Nonzero if the feature is available.
519 feat_debug(const char *ftok
, const char *check
, int verdict
)
522 dispatch_debug("feature `%s': %s -> %s", ftok
, check
,
523 verdict ?
"available" : "absent");
528 int cpu_feature_p(int feat
)
532 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat: \
533 if ((f = feat_debug(ftok, "environment override", \
534 check_env(ftok))) >= 0) \
537 return (feat_debug(ftok, "runtime probe", cond));
540 #if CPUFAM_X86 || CPUFAM_AMD64
541 CASE_CPUFEAT(X86_SSE2
, "x86:sse2",
542 xmm_registers_available_p() &&
543 cpuid_features_p(CPUID1D_SSE2
, 0));
544 CASE_CPUFEAT(X86_AESNI
, "x86:aesni",
545 xmm_registers_available_p() &&
546 cpuid_features_p(CPUID1D_SSE2
, CPUID1C_AESNI
));
547 CASE_CPUFEAT(X86_RDRAND
, "x86:rdrand",
548 cpuid_features_p(0, CPUID1C_RDRAND
));
549 CASE_CPUFEAT(X86_AVX
, "x86:avx",
550 xmm_registers_available_p() &&
551 cpuid_features_p(0, CPUID1C_AVX
));
554 # define FEATP__CASE(feat, tok) \
555 CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
560 dispatch_debug("denying unknown feature %d", feat
);
566 /*----- That's all, folks -------------------------------------------------*/