Home | History | Annotate | Line # | Download | only in crypto
      1 /*
      2  * Copyright 2011-2025 The OpenSSL Project Authors. All Rights Reserved.
      3  *
      4  * Licensed under the Apache License 2.0 (the "License").  You may not use
      5  * this file except in compliance with the License.  You can obtain a copy
      6  * in the file LICENSE in the source distribution or at
      7  * https://www.openssl.org/source/license.html
      8  */
      9 
     10 #include <stdio.h>
     11 #include <stdlib.h>
     12 #include <string.h>
     13 #include <openssl/crypto.h>
     14 #ifdef __APPLE__
     15 #include <sys/sysctl.h>
     16 #else
     17 #include <setjmp.h>
     18 #include <signal.h>
     19 #endif
     20 #include "internal/cryptlib.h"
     21 #ifdef _WIN32
     22 #include <windows.h>
     23 #else
     24 #include <unistd.h>
     25 #endif
     26 #include "arm_arch.h"
     27 
     28 unsigned int OPENSSL_armcap_P = 0;
     29 unsigned int OPENSSL_arm_midr = 0;
     30 unsigned int OPENSSL_armv8_rsa_neonized = 0;
     31 
     32 #ifdef _WIN32
     33 void OPENSSL_cpuid_setup(void)
     34 {
     35     OPENSSL_armcap_P |= ARMV7_NEON;
     36     OPENSSL_armv8_rsa_neonized = 1;
     37     if (IsProcessorFeaturePresent(PF_ARM_V8_CRYPTO_INSTRUCTIONS_AVAILABLE)) {
     38         /* These are all covered by one call in Windows */
     39         OPENSSL_armcap_P |= ARMV8_AES;
     40         OPENSSL_armcap_P |= ARMV8_PMULL;
     41         OPENSSL_armcap_P |= ARMV8_SHA1;
     42         OPENSSL_armcap_P |= ARMV8_SHA256;
     43     }
     44 }
     45 
     46 uint32_t OPENSSL_rdtsc(void)
     47 {
     48     return 0;
     49 }
     50 #elif __ARM_MAX_ARCH__ < 7
     51 void OPENSSL_cpuid_setup(void)
     52 {
     53 }
     54 
     55 uint32_t OPENSSL_rdtsc(void)
     56 {
     57     return 0;
     58 }
     59 #else /* !_WIN32 && __ARM_MAX_ARCH__ >= 7 */
     60 
     61 /* 3 ways of handling things here: __APPLE__,  getauxval() or SIGILL detect */
     62 
     63 /* First determine if getauxval() is available (OSSL_IMPLEMENT_GETAUXVAL) */
     64 
     65 #if defined(__GNUC__) && __GNUC__ >= 2
     66 void OPENSSL_cpuid_setup(void) __attribute__((constructor));
     67 #endif
     68 
     69 #if defined(__GLIBC__) && defined(__GLIBC_PREREQ)
     70 #if __GLIBC_PREREQ(2, 16)
     71 #include <sys/auxv.h>
     72 #define OSSL_IMPLEMENT_GETAUXVAL
     73 #endif
     74 #elif defined(__ANDROID_API__)
     75 /* see https://developer.android.google.cn/ndk/guides/cpu-features */
     76 #if __ANDROID_API__ >= 18
     77 #include <sys/auxv.h>
     78 #define OSSL_IMPLEMENT_GETAUXVAL
     79 #endif
     80 #endif
     81 #if defined(__FreeBSD__) || defined(__OpenBSD__)
     82 #include <sys/param.h>
     83 #if (defined(__FreeBSD__) && __FreeBSD_version >= 1200000) || (defined(__OpenBSD__) && OpenBSD >= 202409)
     84 #include <sys/auxv.h>
     85 #define OSSL_IMPLEMENT_GETAUXVAL
     86 
     87 static unsigned long getauxval(unsigned long key)
     88 {
     89     unsigned long val = 0ul;
     90 
     91     if (elf_aux_info((int)key, &val, sizeof(val)) != 0)
     92         return 0ul;
     93 
     94     return val;
     95 }
     96 #endif
     97 #endif
     98 
     99 /*
    100  * Android: according to https://developer.android.com/ndk/guides/cpu-features,
    101  * getauxval is supported starting with API level 18
    102  */
    103 #if defined(__ANDROID__) && defined(__ANDROID_API__) && __ANDROID_API__ >= 18
    104 #include <sys/auxv.h>
    105 #define OSSL_IMPLEMENT_GETAUXVAL
    106 #endif
    107 
    108 /*
    109  * ARM puts the feature bits for Crypto Extensions in AT_HWCAP2, whereas
    110  * AArch64 used AT_HWCAP.
    111  */
    112 #ifndef AT_HWCAP
    113 #define AT_HWCAP 16
    114 #endif
    115 #ifndef AT_HWCAP2
    116 #define AT_HWCAP2 26
    117 #endif
    118 #if defined(__arm__) || defined(__arm)
    119 #define OSSL_HWCAP AT_HWCAP
    120 #define OSSL_HWCAP_NEON (1 << 12)
    121 
    122 #define OSSL_HWCAP_CE AT_HWCAP2
    123 #define OSSL_HWCAP_CE_AES (1 << 0)
    124 #define OSSL_HWCAP_CE_PMULL (1 << 1)
    125 #define OSSL_HWCAP_CE_SHA1 (1 << 2)
    126 #define OSSL_HWCAP_CE_SHA256 (1 << 3)
    127 #elif defined(__aarch64__)
    128 #define OSSL_HWCAP AT_HWCAP
    129 #define OSSL_HWCAP_NEON (1 << 1)
    130 
    131 #define OSSL_HWCAP_CE AT_HWCAP
    132 #define OSSL_HWCAP_CE_AES (1 << 3)
    133 #define OSSL_HWCAP_CE_PMULL (1 << 4)
    134 #define OSSL_HWCAP_CE_SHA1 (1 << 5)
    135 #define OSSL_HWCAP_CE_SHA256 (1 << 6)
    136 #define OSSL_HWCAP_CPUID (1 << 11)
    137 #define OSSL_HWCAP_SHA3 (1 << 17)
    138 #define OSSL_HWCAP_CE_SM3 (1 << 18)
    139 #define OSSL_HWCAP_CE_SM4 (1 << 19)
    140 #define OSSL_HWCAP_CE_SHA512 (1 << 21)
    141 #define OSSL_HWCAP_SVE (1 << 22)
    142 /* AT_HWCAP2 */
    143 #define OSSL_HWCAP2 26
    144 #define OSSL_HWCAP2_SVE2 (1 << 1)
    145 #define OSSL_HWCAP2_RNG (1 << 16)
    146 #endif
    147 
    148 uint32_t _armv7_tick(void);
    149 
    150 uint32_t OPENSSL_rdtsc(void)
    151 {
    152     if (OPENSSL_armcap_P & ARMV7_TICK)
    153         return _armv7_tick();
    154     else
    155         return 0;
    156 }
    157 
    158 #ifdef __aarch64__
    159 size_t OPENSSL_rndr_asm(unsigned char *buf, size_t len);
    160 size_t OPENSSL_rndrrs_asm(unsigned char *buf, size_t len);
    161 
    162 size_t OPENSSL_rndr_bytes(unsigned char *buf, size_t len);
    163 size_t OPENSSL_rndrrs_bytes(unsigned char *buf, size_t len);
    164 
    165 static size_t OPENSSL_rndr_wrapper(size_t (*func)(unsigned char *, size_t), unsigned char *buf, size_t len)
    166 {
    167     size_t buffer_size = 0;
    168     int i;
    169 
    170     for (i = 0; i < 8; i++) {
    171         buffer_size = func(buf, len);
    172         if (buffer_size == len)
    173             break;
    174         usleep(5000); /* 5000 microseconds (5 milliseconds) */
    175     }
    176     return buffer_size;
    177 }
    178 
    179 size_t OPENSSL_rndr_bytes(unsigned char *buf, size_t len)
    180 {
    181     return OPENSSL_rndr_wrapper(OPENSSL_rndr_asm, buf, len);
    182 }
    183 
    184 size_t OPENSSL_rndrrs_bytes(unsigned char *buf, size_t len)
    185 {
    186     return OPENSSL_rndr_wrapper(OPENSSL_rndrrs_asm, buf, len);
    187 }
    188 #endif
    189 
    190 #if !defined(__APPLE__) && !defined(OSSL_IMPLEMENT_GETAUXVAL)
    191 static sigset_t all_masked;
    192 
    193 static sigjmp_buf ill_jmp;
    194 static void ill_handler(int sig)
    195 {
    196     siglongjmp(ill_jmp, sig);
    197 }
    198 
    199 /*
    200  * Following subroutines could have been inlined, but not all
    201  * ARM compilers support inline assembler, and we'd then have to
    202  * worry about the compiler optimising out the detection code...
    203  */
    204 void _armv7_neon_probe(void);
    205 void _armv8_aes_probe(void);
    206 void _armv8_sha1_probe(void);
    207 void _armv8_sha256_probe(void);
    208 void _armv8_pmull_probe(void);
    209 #ifdef __aarch64__
    210 void _armv8_sm3_probe(void);
    211 void _armv8_sm4_probe(void);
    212 void _armv8_sha512_probe(void);
    213 void _armv8_eor3_probe(void);
    214 void _armv8_sve_probe(void);
    215 void _armv8_sve2_probe(void);
    216 void _armv8_rng_probe(void);
    217 #endif
    218 #endif /* !__APPLE__ && !OSSL_IMPLEMENT_GETAUXVAL */
    219 
    220 /* We only call _armv8_cpuid_probe() if (OPENSSL_armcap_P & ARMV8_CPUID) != 0 */
    221 unsigned int _armv8_cpuid_probe(void);
    222 
    223 #if defined(__APPLE__)
    224 /*
    225  * Checks the specified integer sysctl, returning `value` if it's 1, otherwise returning 0.
    226  */
    227 static unsigned int sysctl_query(const char *name, unsigned int value)
    228 {
    229     unsigned int sys_value = 0;
    230     size_t len = sizeof(sys_value);
    231 
    232     return (sysctlbyname(name, &sys_value, &len, NULL, 0) == 0 && sys_value == 1) ? value : 0;
    233 }
    234 #elif !defined(OSSL_IMPLEMENT_GETAUXVAL)
    235 /*
    236  * Calls a provided probe function, which may SIGILL. If it doesn't, return `value`, otherwise return 0.
    237  */
    238 static unsigned int arm_probe_for(void (*probe)(void), volatile unsigned int value)
    239 {
    240     if (sigsetjmp(ill_jmp, 1) == 0) {
    241         probe();
    242         return value;
    243     } else {
    244         /* The probe function gave us SIGILL */
    245         return 0;
    246     }
    247 }
    248 #endif
    249 
    250 void OPENSSL_cpuid_setup(void)
    251 {
    252     const char *e;
    253 #if !defined(__APPLE__) && !defined(OSSL_IMPLEMENT_GETAUXVAL)
    254     struct sigaction ill_oact, ill_act;
    255     sigset_t oset;
    256 #endif
    257     static int trigger = 0;
    258 
    259     if (trigger)
    260         return;
    261     trigger = 1;
    262 
    263     OPENSSL_armcap_P = 0;
    264 
    265     if ((e = getenv("OPENSSL_armcap"))) {
    266         OPENSSL_armcap_P = (unsigned int)strtoul(e, NULL, 0);
    267         return;
    268     }
    269 
    270 #if defined(__APPLE__)
    271 #if !defined(__aarch64__)
    272     /*
    273      * Capability probing by catching SIGILL appears to be problematic
    274      * on iOS. But since Apple universe is "monocultural", it's actually
    275      * possible to simply set pre-defined processor capability mask.
    276      */
    277     if (1) {
    278         OPENSSL_armcap_P = ARMV7_NEON;
    279         return;
    280     }
    281 #else
    282     {
    283         /*
    284          * From
    285          * https://github.com/llvm/llvm-project/blob/412237dcd07e5a2afbb1767858262a5f037149a3/llvm/lib/Target/AArch64/AArch64.td#L719
    286          * all of these have been available on 64-bit Apple Silicon from the
    287          * beginning (the A7).
    288          */
    289         OPENSSL_armcap_P |= ARMV7_NEON | ARMV8_PMULL | ARMV8_AES | ARMV8_SHA1 | ARMV8_SHA256;
    290 
    291         /* More recent extensions are indicated by sysctls */
    292         OPENSSL_armcap_P |= sysctl_query("hw.optional.armv8_2_sha512", ARMV8_SHA512);
    293         OPENSSL_armcap_P |= sysctl_query("hw.optional.armv8_2_sha3", ARMV8_SHA3);
    294 
    295         if (OPENSSL_armcap_P & ARMV8_SHA3) {
    296             char uarch[64];
    297 
    298             size_t len = sizeof(uarch);
    299             if ((sysctlbyname("machdep.cpu.brand_string", uarch, &len, NULL, 0) == 0) && ((strncmp(uarch, "Apple M1", 8) == 0) || (strncmp(uarch, "Apple M2", 8) == 0) || (strncmp(uarch, "Apple M3", 8) == 0) || (strncmp(uarch, "Apple M4", 8) == 0))) {
    300                 OPENSSL_armcap_P |= ARMV8_UNROLL8_EOR3;
    301                 OPENSSL_armcap_P |= ARMV8_HAVE_SHA3_AND_WORTH_USING;
    302             }
    303         }
    304     }
    305 #endif /* __aarch64__ */
    306 
    307 #elif defined(OSSL_IMPLEMENT_GETAUXVAL)
    308 
    309     if (getauxval(OSSL_HWCAP) & OSSL_HWCAP_NEON) {
    310         unsigned long hwcap = getauxval(OSSL_HWCAP_CE);
    311 
    312         OPENSSL_armcap_P |= ARMV7_NEON;
    313 
    314         if (hwcap & OSSL_HWCAP_CE_AES)
    315             OPENSSL_armcap_P |= ARMV8_AES;
    316 
    317         if (hwcap & OSSL_HWCAP_CE_PMULL)
    318             OPENSSL_armcap_P |= ARMV8_PMULL;
    319 
    320         if (hwcap & OSSL_HWCAP_CE_SHA1)
    321             OPENSSL_armcap_P |= ARMV8_SHA1;
    322 
    323         if (hwcap & OSSL_HWCAP_CE_SHA256)
    324             OPENSSL_armcap_P |= ARMV8_SHA256;
    325 
    326 #ifdef __aarch64__
    327         if (hwcap & OSSL_HWCAP_CE_SM4)
    328             OPENSSL_armcap_P |= ARMV8_SM4;
    329 
    330         if (hwcap & OSSL_HWCAP_CE_SHA512)
    331             OPENSSL_armcap_P |= ARMV8_SHA512;
    332 
    333         if (hwcap & OSSL_HWCAP_CPUID)
    334             OPENSSL_armcap_P |= ARMV8_CPUID;
    335 
    336         if (hwcap & OSSL_HWCAP_CE_SM3)
    337             OPENSSL_armcap_P |= ARMV8_SM3;
    338         if (hwcap & OSSL_HWCAP_SHA3)
    339             OPENSSL_armcap_P |= ARMV8_SHA3;
    340 #endif
    341     }
    342 #ifdef __aarch64__
    343     if (getauxval(OSSL_HWCAP) & OSSL_HWCAP_SVE)
    344         OPENSSL_armcap_P |= ARMV8_SVE;
    345 
    346     if (getauxval(OSSL_HWCAP2) & OSSL_HWCAP2_SVE2)
    347         OPENSSL_armcap_P |= ARMV8_SVE2;
    348 
    349     if (getauxval(OSSL_HWCAP2) & OSSL_HWCAP2_RNG)
    350         OPENSSL_armcap_P |= ARMV8_RNG;
    351 #endif
    352 
    353 #else /* !__APPLE__ && !OSSL_IMPLEMENT_GETAUXVAL */
    354 
    355     /* If all else fails, do brute force SIGILL-based feature detection */
    356 
    357     sigfillset(&all_masked);
    358     sigdelset(&all_masked, SIGILL);
    359     sigdelset(&all_masked, SIGTRAP);
    360     sigdelset(&all_masked, SIGFPE);
    361     sigdelset(&all_masked, SIGBUS);
    362     sigdelset(&all_masked, SIGSEGV);
    363 
    364     memset(&ill_act, 0, sizeof(ill_act));
    365     ill_act.sa_handler = ill_handler;
    366     ill_act.sa_mask = all_masked;
    367 
    368     sigprocmask(SIG_SETMASK, &ill_act.sa_mask, &oset);
    369     sigaction(SIGILL, &ill_act, &ill_oact);
    370 
    371     OPENSSL_armcap_P |= arm_probe_for(_armv7_neon_probe, ARMV7_NEON);
    372 
    373     if (OPENSSL_armcap_P & ARMV7_NEON) {
    374 
    375         OPENSSL_armcap_P |= arm_probe_for(_armv8_pmull_probe, ARMV8_PMULL | ARMV8_AES);
    376         if (!(OPENSSL_armcap_P & ARMV8_AES)) {
    377             OPENSSL_armcap_P |= arm_probe_for(_armv8_aes_probe, ARMV8_AES);
    378         }
    379 
    380         OPENSSL_armcap_P |= arm_probe_for(_armv8_sha1_probe, ARMV8_SHA1);
    381         OPENSSL_armcap_P |= arm_probe_for(_armv8_sha256_probe, ARMV8_SHA256);
    382 
    383 #if defined(__aarch64__)
    384         OPENSSL_armcap_P |= arm_probe_for(_armv8_sm3_probe, ARMV8_SM3);
    385         OPENSSL_armcap_P |= arm_probe_for(_armv8_sm4_probe, ARMV8_SM4);
    386         OPENSSL_armcap_P |= arm_probe_for(_armv8_sha512_probe, ARMV8_SHA512);
    387         OPENSSL_armcap_P |= arm_probe_for(_armv8_eor3_probe, ARMV8_SHA3);
    388 #endif
    389     }
    390 #ifdef __aarch64__
    391     OPENSSL_armcap_P |= arm_probe_for(_armv8_sve_probe, ARMV8_SVE);
    392     OPENSSL_armcap_P |= arm_probe_for(_armv8_sve2_probe, ARMV8_SVE2);
    393     OPENSSL_armcap_P |= arm_probe_for(_armv8_rng_probe, ARMV8_RNG);
    394 #endif
    395 
    396     /*
    397      * Probing for ARMV7_TICK is known to produce unreliable results,
    398      * so we only use the feature when the user explicitly enables it
    399      * with OPENSSL_armcap.
    400      */
    401 
    402     sigaction(SIGILL, &ill_oact, NULL);
    403     sigprocmask(SIG_SETMASK, &oset, NULL);
    404 
    405 #endif /* __APPLE__, OSSL_IMPLEMENT_GETAUXVAL */
    406 
    407 #ifdef __aarch64__
    408     if (OPENSSL_armcap_P & ARMV8_CPUID)
    409         OPENSSL_arm_midr = _armv8_cpuid_probe();
    410 
    411     if ((MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_ARM, ARM_CPU_PART_CORTEX_A72) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_ARM, ARM_CPU_PART_N1)) && (OPENSSL_armcap_P & ARMV7_NEON)) {
    412         OPENSSL_armv8_rsa_neonized = 1;
    413     }
    414     if ((MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_ARM, ARM_CPU_PART_V1) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_ARM, ARM_CPU_PART_N2) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_MICROSOFT, MICROSOFT_CPU_PART_COBALT_100) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_ARM, ARM_CPU_PART_V2) || MIDR_IMPLEMENTER(OPENSSL_arm_midr) == ARM_CPU_IMP_AMPERE) && (OPENSSL_armcap_P & ARMV8_SHA3))
    415         OPENSSL_armcap_P |= ARMV8_UNROLL8_EOR3;
    416     if ((MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_ARM, ARM_CPU_PART_V1) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_ARM, ARM_CPU_PART_V2) || MIDR_IMPLEMENTER(OPENSSL_arm_midr) == ARM_CPU_IMP_AMPERE) && (OPENSSL_armcap_P & ARMV8_SHA3))
    417         OPENSSL_armcap_P |= ARMV8_UNROLL12_EOR3;
    418     if ((MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M1_FIRESTORM) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M1_ICESTORM) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M1_FIRESTORM_PRO) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M1_ICESTORM_PRO) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M1_FIRESTORM_MAX) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M1_ICESTORM_MAX) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M2_AVALANCHE) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M2_BLIZZARD) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M2_AVALANCHE_PRO) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M2_BLIZZARD_PRO) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M2_AVALANCHE_MAX) || MIDR_IS_CPU_MODEL(OPENSSL_arm_midr, ARM_CPU_IMP_APPLE, APPLE_CPU_PART_M2_BLIZZARD_MAX)) && (OPENSSL_armcap_P & ARMV8_SHA3))
    419         OPENSSL_armcap_P |= ARMV8_HAVE_SHA3_AND_WORTH_USING;
    420 #endif
    421 }
    422 #endif /* _WIN32, __ARM_MAX_ARCH__ >= 7 */
    423