459302f0913202673bb46d6dba854fc0b6d4782c
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 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)
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);
58 static void cpuid(struct cpuid
*cc
, unsigned a
, unsigned c
)
60 int rc
= dispatch_x86ish_cpuid(cc
, a
, c
);
62 dispatch_debug("CPUID instruction not available");
64 dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x",
65 a
, c
, cc
->a
, cc
->b
, cc
->c
, cc
->d
);
68 static unsigned cpuid_maxleaf(void)
69 { struct cpuid c
; cpuid(&c
, 0, 0); return (c
.a
); }
71 /* --- @cpuid_features_p@ --- *
73 * Arguments: @unsigned dbits@ = bits to check in EDX
74 * @unsigned cbits@ = bits to check in ECX
76 * Returns: Nonzero if all the requested bits are set in the CPUID result
80 static int cpuid_features_p(unsigned dbits
, unsigned cbits
)
83 if (cpuid_maxleaf() < 1) return (0);
85 return ((c
.d
& dbits
) == dbits
&& (c
.c
& cbits
) == cbits
);
88 /* --- @xmm_registers_available_p@ --- *
92 * Returns: Nonzero if the operating system has made the XMM registers
96 static int xmm_registers_available_p(void)
98 int f
= dispatch_x86ish_xmmregisters_p();
100 dispatch_debug("XMM registers %savailable", f ?
"" : "not ");
104 /* --- @rdrand_works_p@ --- *
109 * Returns: Nonzero if the `rdrand' instruction actually works. Assumes
110 * that it's already been verified to be safe to issue.
114 static int rdrand(unsigned *x
)
120 __asm__ ("" : "=g" (_t
));
121 __asm__ ("0: rdrand %2; jc 1f; decl %1; jnz 0b\n"
122 "mov $-1, %0; jmp 9f\n"
123 "1: movl %2, (%3); xorl %0, %0\n"
125 : "=r" (rc
), "+r" (i
), "+r" (_t
)
132 static int rdrand_works_p(void)
136 /* Check that it doesn't always give the same answer. Try four times: this
137 * will fail with probability %$2^{-128}$% with a truly random generator,
138 * which seems fair enough.
140 if (rdrand(&ref
)) goto fail
;
141 for (i
= 0; i
< 4; i
++) {
142 if (rdrand(&x
)) goto fail
;
143 if (x
!= ref
) goto not_stuck
;
145 dispatch_debug("RDRAND always returns 0x%08x!", ref
);
149 dispatch_debug("RDRAND instruction looks plausible");
153 dispatch_debug("RDRAND instruction fails too often");
159 /*----- General feature probing using auxiliary vectors -------------------*/
161 /* Try to find the system's definitions for auxiliary vector entries. */
162 #ifdef HAVE_SYS_AUXV_H
163 # include <sys/auxv.h>
165 #ifdef HAVE_LINUX_AUXVEC_H
166 # include <linux/auxvec.h>
168 #ifdef HAVE_ASM_HWCAP_H
169 # include <asm/hwcap.h>
172 /* The type of entries in the auxiliary vector. I'm assuming that `unsigned
173 * long' matches each platform's word length; if this is false then we'll
174 * need some host-specific tweaking here.
176 union auxval
{ long i
; unsigned long u
; const void *p
; };
177 struct auxentry
{ unsigned long type
; union auxval value
; };
179 /* Register each CPU family's interest in the auxiliary vector. Make sure
180 * that the necessary entry types are defined. This is primarily ordered by
181 * entry type to minimize duplication.
183 #if defined(AT_HWCAP) && CPUFAM_ARMEL
185 # define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
188 #if defined(AT_HWCAP) && CPUFAM_ARM64
190 # define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
193 #if defined(AT_HWCAP2) && CPUFAM_ARMEL
195 # define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
198 /* If we couldn't find any interesting entries then we can switch all of this
199 * machinery off. Also do that if we have no means for atomic updates.
201 #if WANT_ANY && CPU_DISPATCH_P
203 /* The main output of this section is a bitmask of detected features. The
204 * least significant bit will be set if we've tried to probe. Always access
205 * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
207 static unsigned hwcaps
= 0;
209 /* For each potentially interesting type which turned out not to exist or be
210 * wanted, define a dummy macro for the sake of the next step.
212 #ifndef WANT_AT_HWCAP
213 # define WANT_AT_HWCAP(_)
215 #ifndef WANT_AT_HWCAP2
216 # define WANT_AT_HWCAP2(_)
219 /* For each CPU family, define two lists.
221 * * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
222 * family tried to register interest in above. Each entry contains the
223 * interesting auxiliary vector entry type, the name of the union branch
224 * for its value, and the name of the slot in `struct auxprobe' in which
225 * to store the value.
227 * * `CAPMAP' is a list describing the output features which the CPU family
228 * intends to satisfy from the auxiliary vector. Each entry contains a
229 * feature name suffix, and the token name (for `check_env').
232 # define WANTAUX(_) \
236 _(ARM_VFP, "arm:vfp") \
237 _(ARM_NEON, "arm:neon") \
238 _(ARM_V4, "arm:v4") \
239 _(ARM_D32, "arm:d32") \
240 _(ARM_AES, "arm:aes") \
241 _(ARM_PMULL, "arm:pmull")
244 # define WANTAUX(_) \
247 _(ARM_AES, "arm:aes") \
248 _(ARM_PMULL, "arm:pmull")
251 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
254 #define HFI__ENUM(feat, tok) HFI_##feat,
261 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
267 /* Build a structure in which we can capture the interesting data from the
270 #define AUXUTYPE_i long
271 #define AUXUTYPE_u unsigned long
272 #define AUXUTYPE_p const void *
274 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
275 WANTAUX(AUXPROBE__SLOT
)
279 /* --- @probe_hwcaps@ --- *
285 * Use: Attempt to find the auxiliary vector (which is well hidden)
286 * and discover interesting features from it.
289 static void probe_hwcaps(void)
291 unsigned hw
= HF_PROBED
;
292 struct auxprobe probed
= { 0 };
294 /* Populate `probed' with the information we manage to retrieve from the
295 * auxiliary vector. Slots we couldn't find are left zero-valued.
297 #if defined(HAVE_GETAUXVAL)
298 /* Shiny new libc lets us request individual entry types. This is almost
301 # define CAP__GET(type, ubranch, slot) \
302 probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
305 /* Otherwise we're a bit stuck, really. Modern Linux kernels make a copy
306 * of the vector available in `/procc' so we could try that.
308 * The usual place is stuck on the end of the environment vector, but that
309 * may well have moved, and we have no way of telling whether it has or
310 * whether there was ever an auxiliary vector there at all; so don't do
315 unsigned char *p
= 0, *q
= 0;
316 const struct auxentry
*a
;
319 /* Open the file and read it into a memory chunk. */
320 if ((fp
= fopen("/proc/self/auxv", "rb")) == 0) goto clean
;
322 if ((p
= malloc(sz
)) == 0) goto clean
;
324 n
= fread(p
+ off
, 1, sz
- off
, fp
);
327 sz
*= 2; if ((q
= realloc(p
, sz
)) == 0) break;
331 /* Work through the vector (or as much of it as we found) and extract the
332 * types we're interested in.
334 for (a
= (const struct auxentry
*)p
,
335 n
= sz
/sizeof(struct auxentry
);
338 #define CAP__SWITCH(type, ubranch, slot) \
339 case type: probed.slot = a->value.ubranch; break;
341 case AT_NULL
: goto clean
;
351 /* Each CPU family now has to pick through what was found and stashed in
352 * `probed', and set the appropriate flag bits in `hw'.
355 if (probed
.hwcap
& HWCAP_VFPv3
) hw
|= HF_ARM_VFP
;
356 if (probed
.hwcap
& HWCAP_NEON
) hw
|= HF_ARM_NEON
;
357 if (probed
.hwcap
& HWCAP_VFPD32
) hw
|= HF_ARM_D32
;
358 if (probed
.hwcap
& HWCAP_VFPv4
) hw
|= HF_ARM_V4
;
360 if (probed
.hwcap2
& HWCAP2_AES
) hw
|= HF_ARM_AES
;
363 if (probed
.hwcap2
& HWCAP2_PMULL
) hw
|= HF_ARM_PMULL
;
367 if (probed
.hwcap
& HWCAP_AES
) hw
|= HF_ARM_AES
;
368 if (probed
.hwcap
& HWCAP_PMULL
) hw
|= HF_ARM_PMULL
;
371 /* Store the bitmask of features we probed for everyone to see. */
372 DISPATCH_STORE(hwcaps
, hw
);
374 /* Finally, make a report about the things we found. (Doing this earlier
375 * will pointlessly widen the window in which multiple threads will do the
376 * above auxiliary-vector probing.)
378 #define CAP__DEBUG(feat, tok) \
379 dispatch_debug("check auxv for feature `%s': %s", tok, \
380 hw & HF_##feat ? "available" : "absent");
385 /* --- @get_hwcaps@ --- *
389 * Returns: A mask of hardware capabilities and other features, as probed
390 * from the auxiliary vector.
393 static unsigned get_hwcaps(void)
397 DISPATCH_LOAD(hwcaps
, hw
);
398 if (!(hwcaps
& HF_PROBED
)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps
, hw
); }
404 /*----- External interface ------------------------------------------------*/
406 /* --- @dispatch_debug@ --- *
408 * Arguments: @const char *fmt@ = a format string
409 * @...@ = additional arguments
413 * Use: Writes a formatted message to standard output if dispatch
414 * debugging is enabled.
417 void dispatch_debug(const char *fmt
, ...)
420 const char *e
= getenv("CATACOMB_CPUDISPATCH_DEBUG");
422 if (e
&& *e
!= 'n' && *e
!= '0') {
424 fputs("Catacomb CPUDISPATCH: ", stderr
);
425 vfprintf(stderr
, fmt
, ap
);
431 /* --- @check_env@ --- *
433 * Arguments: @const char *ftok@ = feature token
435 * Returns: Zero if the feature is forced off; positive if it's forced
436 * on; negative if the user hasn't decided.
438 * Use: Checks the environment variable `CATACOMB_CPUFEAT' for the
439 * feature token @ftok@. The variable, if it exists, should be
440 * a space-separated sequence of `+tok' and `-tok' items. These
441 * tokens may end in `*', which matches any suffix.
444 static int IGNORABLE
check_env(const char *ftok
)
446 const char *p
, *q
, *pp
;
449 p
= getenv("CATACOMB_CPUFEAT");
453 while (isspace((unsigned char)*p
)) p
++;
454 if (!*p
) return (-1);
456 case '+': d
= +1; p
++; break;
457 case '-': d
= 0; p
++; break;
458 default: d
= -1; break;
460 for (q
= p
; *q
&& !isspace((unsigned char)*q
); q
++);
462 for (pp
= ftok
; p
< q
&& *pp
&& *p
== *pp
; p
++, pp
++);
463 if ((p
== q
&& !*pp
) || (*p
== '*' && p
+ 1 == q
)) return (d
);
470 /* --- @cpu_feature_p@ --- *
472 * Arguments: @unsigned feat@ = a @CPUFEAT_...@ code
474 * Returns: Nonzero if the feature is available.
480 feat_debug(const char *ftok
, const char *check
, int verdict
)
483 dispatch_debug("feature `%s': %s -> %s", ftok
, check
,
484 verdict ?
"available" : "absent");
489 int cpu_feature_p(int feat
)
493 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat: \
494 if ((f = feat_debug(ftok, "environment override", \
495 check_env(ftok))) >= 0) \
498 return (feat_debug(ftok, "runtime probe", cond));
501 #if CPUFAM_X86 || CPUFAM_AMD64
502 CASE_CPUFEAT(X86_SSE2
, "x86:sse2",
503 cpuid_features_p(CPUID1D_SSE2
, 0) &&
504 xmm_registers_available_p());
505 CASE_CPUFEAT(X86_AESNI
, "x86:aesni",
506 cpuid_features_p(CPUID1D_SSE2
, CPUID1C_AESNI
) &&
507 xmm_registers_available_p());
508 CASE_CPUFEAT(X86_RDRAND
, "x86:rdrand",
509 cpuid_features_p(0, CPUID1C_RDRAND
) && rdrand_works_p());
510 CASE_CPUFEAT(X86_AVX
, "x86:avx",
511 cpuid_features_p(0, CPUID1C_AVX
) &&
512 xmm_registers_available_p());
513 CASE_CPUFEAT(X86_SSSE3
, "x86:ssse3",
514 cpuid_features_p(0, CPUID1C_SSSE3
) &&
515 xmm_registers_available_p());
516 CASE_CPUFEAT(X86_PCLMUL
, "x86:pclmul",
517 cpuid_features_p(0, CPUID1C_PCLMUL
) &&
518 xmm_registers_available_p());
521 # define FEATP__CASE(feat, tok) \
522 CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
527 dispatch_debug("denying unknown feature %d", feat
);
533 /*----- That's all, folks -------------------------------------------------*/