]>
Commit | Line | Data |
---|---|---|
b1322259 RS |
1 | /* |
2 | * Copyright 2009-2016 The OpenSSL Project Authors. All Rights Reserved. | |
3 | * | |
4 | * Licensed under the OpenSSL license (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 | ||
b4b48a10 AP |
10 | #include <stdio.h> |
11 | #include <stdlib.h> | |
12 | #include <string.h> | |
13 | #include <setjmp.h> | |
14 | #include <signal.h> | |
fd054957 | 15 | #include <unistd.h> |
d5630dd6 | 16 | #if defined(__linux) || defined(_AIX) |
0f113f3e | 17 | # include <sys/utsname.h> |
78c3e205 | 18 | #endif |
2688d999 AP |
19 | #if defined(_AIX53) /* defined even on post-5.3 */ |
20 | # include <sys/systemcfg.h> | |
21 | # if !defined(__power_set) | |
22 | # define __power_set(a) (_system_configuration.implementation & (a)) | |
23 | # endif | |
24 | #endif | |
0bd93bbe AP |
25 | #if defined(__APPLE__) && defined(__MACH__) |
26 | # include <sys/types.h> | |
27 | # include <sys/sysctl.h> | |
28 | #endif | |
11252459 | 29 | #include <openssl/crypto.h> |
10b0b5ec | 30 | #include <openssl/bn.h> |
b4b48a10 | 31 | |
0e716d92 | 32 | #include "ppc_arch.h" |
b4b48a10 | 33 | |
07f3e4f3 | 34 | unsigned int OPENSSL_ppccap_P = 0; |
b4b48a10 AP |
35 | |
36 | static sigset_t all_masked; | |
37 | ||
10b0b5ec RL |
38 | #ifdef OPENSSL_BN_ASM_MONT |
39 | int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, | |
40 | const BN_ULONG *np, const BN_ULONG *n0, int num) | |
41 | { | |
42 | int bn_mul_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, | |
43 | const BN_ULONG *np, const BN_ULONG *n0, int num); | |
44 | int bn_mul4x_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, | |
45 | const BN_ULONG *np, const BN_ULONG *n0, int num); | |
46 | ||
47 | if (num < 4) | |
48 | return 0; | |
49 | ||
50 | if ((num & 3) == 0) | |
51 | return bn_mul4x_mont_int(rp, ap, bp, np, n0, num); | |
52 | ||
53 | /* | |
54 | * There used to be [optional] call to bn_mul_mont_fpu64 here, | |
55 | * but above subroutine is faster on contemporary processors. | |
56 | * Formulation means that there might be old processors where | |
57 | * FPU code path would be faster, POWER6 perhaps, but there was | |
58 | * no opportunity to figure it out... | |
59 | */ | |
60 | ||
61 | return bn_mul_mont_int(rp, ap, bp, np, n0, num); | |
62 | } | |
63 | #endif | |
64 | ||
65 | void sha256_block_p8(void *ctx, const void *inp, size_t len); | |
66 | void sha256_block_ppc(void *ctx, const void *inp, size_t len); | |
67 | void sha256_block_data_order(void *ctx, const void *inp, size_t len) | |
68 | { | |
69 | OPENSSL_ppccap_P & PPC_CRYPTO207 ? sha256_block_p8(ctx, inp, len) : | |
70 | sha256_block_ppc(ctx, inp, len); | |
71 | } | |
72 | ||
73 | void sha512_block_p8(void *ctx, const void *inp, size_t len); | |
74 | void sha512_block_ppc(void *ctx, const void *inp, size_t len); | |
75 | void sha512_block_data_order(void *ctx, const void *inp, size_t len) | |
76 | { | |
77 | OPENSSL_ppccap_P & PPC_CRYPTO207 ? sha512_block_p8(ctx, inp, len) : | |
78 | sha512_block_ppc(ctx, inp, len); | |
79 | } | |
80 | ||
81 | #ifndef OPENSSL_NO_CHACHA | |
82 | void ChaCha20_ctr32_int(unsigned char *out, const unsigned char *inp, | |
83 | size_t len, const unsigned int key[8], | |
84 | const unsigned int counter[4]); | |
85 | void ChaCha20_ctr32_vmx(unsigned char *out, const unsigned char *inp, | |
86 | size_t len, const unsigned int key[8], | |
87 | const unsigned int counter[4]); | |
88 | void ChaCha20_ctr32(unsigned char *out, const unsigned char *inp, | |
89 | size_t len, const unsigned int key[8], | |
90 | const unsigned int counter[4]) | |
91 | { | |
92 | OPENSSL_ppccap_P & PPC_ALTIVEC | |
93 | ? ChaCha20_ctr32_vmx(out, inp, len, key, counter) | |
94 | : ChaCha20_ctr32_int(out, inp, len, key, counter); | |
95 | } | |
96 | #endif | |
97 | ||
98 | #ifndef OPENSSL_NO_POLY1305 | |
99 | void poly1305_init_int(void *ctx, const unsigned char key[16]); | |
100 | void poly1305_blocks(void *ctx, const unsigned char *inp, size_t len, | |
101 | unsigned int padbit); | |
102 | void poly1305_emit(void *ctx, unsigned char mac[16], | |
103 | const unsigned int nonce[4]); | |
104 | void poly1305_init_fpu(void *ctx, const unsigned char key[16]); | |
105 | void poly1305_blocks_fpu(void *ctx, const unsigned char *inp, size_t len, | |
106 | unsigned int padbit); | |
107 | void poly1305_emit_fpu(void *ctx, unsigned char mac[16], | |
108 | const unsigned int nonce[4]); | |
109 | int poly1305_init(void *ctx, const unsigned char key[16], void *func[2]) | |
110 | { | |
111 | if (sizeof(size_t) == 4 && (OPENSSL_ppccap_P & PPC_FPU)) { | |
112 | poly1305_init_fpu(ctx, key); | |
113 | func[0] = poly1305_blocks_fpu; | |
114 | func[1] = poly1305_emit_fpu; | |
115 | } else { | |
116 | poly1305_init_int(ctx, key); | |
117 | func[0] = poly1305_blocks; | |
118 | func[1] = poly1305_emit; | |
119 | } | |
120 | return 1; | |
121 | } | |
122 | #endif | |
123 | ||
124 | #ifdef ECP_NISTZ256_ASM | |
125 | void ecp_nistz256_mul_mont(unsigned long res[4], const unsigned long a[4], | |
126 | const unsigned long b[4]); | |
127 | ||
128 | void ecp_nistz256_to_mont(unsigned long res[4], const unsigned long in[4]); | |
129 | void ecp_nistz256_to_mont(unsigned long res[4], const unsigned long in[4]) | |
130 | { | |
131 | static const unsigned long RR[] = { 0x0000000000000003U, | |
132 | 0xfffffffbffffffffU, | |
133 | 0xfffffffffffffffeU, | |
134 | 0x00000004fffffffdU }; | |
135 | ||
136 | ecp_nistz256_mul_mont(res, in, RR); | |
137 | } | |
138 | ||
139 | void ecp_nistz256_from_mont(unsigned long res[4], const unsigned long in[4]); | |
140 | void ecp_nistz256_from_mont(unsigned long res[4], const unsigned long in[4]) | |
141 | { | |
142 | static const unsigned long one[] = { 1, 0, 0, 0 }; | |
143 | ||
144 | ecp_nistz256_mul_mont(res, in, one); | |
145 | } | |
146 | #endif | |
147 | ||
b4b48a10 | 148 | static sigjmp_buf ill_jmp; |
0f113f3e MC |
149 | static void ill_handler(int sig) |
150 | { | |
151 | siglongjmp(ill_jmp, sig); | |
152 | } | |
b4b48a10 | 153 | |
2688d999 | 154 | void OPENSSL_fpu_probe(void); |
70b76d39 | 155 | void OPENSSL_ppc64_probe(void); |
fd054957 | 156 | void OPENSSL_altivec_probe(void); |
de51e830 | 157 | void OPENSSL_crypto207_probe(void); |
53385e1f | 158 | void OPENSSL_madd300_probe(void); |
70b76d39 | 159 | |
2688d999 AP |
160 | /* |
161 | * Use a weak reference to getauxval() so we can use it if it is available | |
162 | * but don't break the build if it is not. Note that this is *link-time* | |
163 | * feature detection, not *run-time*. In other words if we link with | |
164 | * symbol present, it's expected to be present even at run-time. | |
165 | */ | |
166 | #if defined(__GNUC__) && __GNUC__>=2 && defined(__ELF__) | |
167 | extern unsigned long getauxval(unsigned long type) __attribute__ ((weak)); | |
168 | #else | |
169 | static unsigned long (*getauxval) (unsigned long) = NULL; | |
170 | #endif | |
171 | ||
172 | /* I wish <sys/auxv.h> was universally available */ | |
173 | #define HWCAP 16 /* AT_HWCAP */ | |
174 | #define HWCAP_PPC64 (1U << 30) | |
175 | #define HWCAP_ALTIVEC (1U << 28) | |
176 | #define HWCAP_FPU (1U << 27) | |
177 | #define HWCAP_POWER6_EXT (1U << 9) | |
178 | #define HWCAP_VSX (1U << 7) | |
179 | ||
180 | #define HWCAP2 26 /* AT_HWCAP2 */ | |
181 | #define HWCAP_VEC_CRYPTO (1U << 25) | |
e0e53282 | 182 | #define HWCAP_ARCH_3_00 (1U << 23) |
2688d999 AP |
183 | |
184 | # if defined(__GNUC__) && __GNUC__>=2 | |
185 | __attribute__ ((constructor)) | |
186 | # endif | |
b4b48a10 | 187 | void OPENSSL_cpuid_setup(void) |
0f113f3e MC |
188 | { |
189 | char *e; | |
190 | struct sigaction ill_oact, ill_act; | |
191 | sigset_t oset; | |
192 | static int trigger = 0; | |
193 | ||
194 | if (trigger) | |
195 | return; | |
196 | trigger = 1; | |
197 | ||
0f113f3e MC |
198 | if ((e = getenv("OPENSSL_ppccap"))) { |
199 | OPENSSL_ppccap_P = strtoul(e, NULL, 0); | |
200 | return; | |
201 | } | |
b4b48a10 | 202 | |
0f113f3e | 203 | OPENSSL_ppccap_P = 0; |
6415dd7b | 204 | |
fd054957 | 205 | #if defined(_AIX) |
2688d999 AP |
206 | OPENSSL_ppccap_P |= PPC_FPU; |
207 | ||
0f113f3e MC |
208 | if (sizeof(size_t) == 4) { |
209 | struct utsname uts; | |
fd054957 | 210 | # if defined(_SC_AIX_KERNEL_BITMODE) |
0f113f3e MC |
211 | if (sysconf(_SC_AIX_KERNEL_BITMODE) != 64) |
212 | return; | |
fd054957 | 213 | # endif |
0f113f3e MC |
214 | if (uname(&uts) != 0 || atoi(uts.version) < 6) |
215 | return; | |
216 | } | |
2688d999 AP |
217 | |
218 | # if defined(__power_set) | |
219 | /* | |
220 | * Value used in __power_set is a single-bit 1<<n one denoting | |
221 | * specific processor class. Incidentally 0xffffffff<<n can be | |
222 | * used to denote specific processor and its successors. | |
223 | */ | |
224 | if (sizeof(size_t) == 4) { | |
225 | /* In 32-bit case PPC_FPU64 is always fastest [if option] */ | |
226 | if (__power_set(0xffffffffU<<13)) /* POWER5 and later */ | |
227 | OPENSSL_ppccap_P |= PPC_FPU64; | |
228 | } else { | |
229 | /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */ | |
230 | if (__power_set(0x1U<<14)) /* POWER6 */ | |
231 | OPENSSL_ppccap_P |= PPC_FPU64; | |
232 | } | |
233 | ||
234 | if (__power_set(0xffffffffU<<14)) /* POWER6 and later */ | |
235 | OPENSSL_ppccap_P |= PPC_ALTIVEC; | |
236 | ||
237 | if (__power_set(0xffffffffU<<16)) /* POWER8 and later */ | |
238 | OPENSSL_ppccap_P |= PPC_CRYPTO207; | |
239 | ||
e0e53282 AP |
240 | if (__power_set(0xffffffffU<<17)) /* POWER9 and later */ |
241 | OPENSSL_ppccap_P |= PPC_MADD300; | |
242 | ||
2688d999 AP |
243 | return; |
244 | # endif | |
245 | #endif | |
246 | ||
0bd93bbe AP |
247 | #if defined(__APPLE__) && defined(__MACH__) |
248 | OPENSSL_ppccap_P |= PPC_FPU; | |
249 | ||
250 | { | |
251 | int val; | |
252 | size_t len = sizeof(val); | |
253 | ||
254 | if (sysctlbyname("hw.optional.64bitops", &val, &len, NULL, 0) == 0) { | |
255 | if (val) | |
256 | OPENSSL_ppccap_P |= PPC_FPU64; | |
257 | } | |
258 | ||
259 | len = sizeof(val); | |
260 | if (sysctlbyname("hw.optional.altivec", &val, &len, NULL, 0) == 0) { | |
261 | if (val) | |
262 | OPENSSL_ppccap_P |= PPC_ALTIVEC; | |
263 | } | |
264 | ||
265 | return; | |
266 | } | |
267 | #endif | |
268 | ||
2688d999 AP |
269 | if (getauxval != NULL) { |
270 | unsigned long hwcap = getauxval(HWCAP); | |
271 | ||
272 | if (hwcap & HWCAP_FPU) { | |
dccd20d1 | 273 | OPENSSL_ppccap_P |= PPC_FPU; |
2688d999 AP |
274 | |
275 | if (sizeof(size_t) == 4) { | |
276 | /* In 32-bit case PPC_FPU64 is always fastest [if option] */ | |
277 | if (hwcap & HWCAP_PPC64) | |
278 | OPENSSL_ppccap_P |= PPC_FPU64; | |
279 | } else { | |
280 | /* In 64-bit case PPC_FPU64 is fastest only on POWER6 */ | |
281 | if (hwcap & HWCAP_POWER6_EXT) | |
282 | OPENSSL_ppccap_P |= PPC_FPU64; | |
283 | } | |
284 | } | |
285 | ||
286 | if (hwcap & HWCAP_ALTIVEC) { | |
287 | OPENSSL_ppccap_P |= PPC_ALTIVEC; | |
288 | ||
289 | if ((hwcap & HWCAP_VSX) && (getauxval(HWCAP2) & HWCAP_VEC_CRYPTO)) | |
290 | OPENSSL_ppccap_P |= PPC_CRYPTO207; | |
291 | } | |
292 | ||
e0e53282 AP |
293 | if (hwcap & HWCAP_ARCH_3_00) { |
294 | OPENSSL_ppccap_P |= PPC_MADD300; | |
295 | } | |
296 | ||
2688d999 AP |
297 | return; |
298 | } | |
299 | ||
300 | sigfillset(&all_masked); | |
301 | sigdelset(&all_masked, SIGILL); | |
302 | sigdelset(&all_masked, SIGTRAP); | |
303 | #ifdef SIGEMT | |
304 | sigdelset(&all_masked, SIGEMT); | |
fd054957 | 305 | #endif |
2688d999 AP |
306 | sigdelset(&all_masked, SIGFPE); |
307 | sigdelset(&all_masked, SIGBUS); | |
308 | sigdelset(&all_masked, SIGSEGV); | |
fd054957 | 309 | |
0f113f3e MC |
310 | memset(&ill_act, 0, sizeof(ill_act)); |
311 | ill_act.sa_handler = ill_handler; | |
312 | ill_act.sa_mask = all_masked; | |
6415dd7b | 313 | |
0f113f3e MC |
314 | sigprocmask(SIG_SETMASK, &ill_act.sa_mask, &oset); |
315 | sigaction(SIGILL, &ill_act, &ill_oact); | |
6415dd7b | 316 | |
2688d999 AP |
317 | if (sigsetjmp(ill_jmp,1) == 0) { |
318 | OPENSSL_fpu_probe(); | |
319 | OPENSSL_ppccap_P |= PPC_FPU; | |
320 | ||
321 | if (sizeof(size_t) == 4) { | |
78c3e205 | 322 | #ifdef __linux |
2688d999 AP |
323 | struct utsname uts; |
324 | if (uname(&uts) == 0 && strcmp(uts.machine, "ppc64") == 0) | |
78c3e205 | 325 | #endif |
2688d999 AP |
326 | if (sigsetjmp(ill_jmp, 1) == 0) { |
327 | OPENSSL_ppc64_probe(); | |
328 | OPENSSL_ppccap_P |= PPC_FPU64; | |
329 | } | |
330 | } else { | |
331 | /* | |
332 | * Wanted code detecting POWER6 CPU and setting PPC_FPU64 | |
333 | */ | |
334 | } | |
0f113f3e MC |
335 | } |
336 | ||
337 | if (sigsetjmp(ill_jmp, 1) == 0) { | |
338 | OPENSSL_altivec_probe(); | |
339 | OPENSSL_ppccap_P |= PPC_ALTIVEC; | |
340 | if (sigsetjmp(ill_jmp, 1) == 0) { | |
341 | OPENSSL_crypto207_probe(); | |
342 | OPENSSL_ppccap_P |= PPC_CRYPTO207; | |
343 | } | |
344 | } | |
345 | ||
e0e53282 AP |
346 | if (sigsetjmp(ill_jmp, 1) == 0) { |
347 | OPENSSL_madd300_probe(); | |
348 | OPENSSL_ppccap_P |= PPC_MADD300; | |
349 | } | |
350 | ||
0f113f3e MC |
351 | sigaction(SIGILL, &ill_oact, NULL); |
352 | sigprocmask(SIG_SETMASK, &oset, NULL); | |
353 | } |