1 | /* Copyright (C) 1995-1998 Eric Young (eay@cryptsoft.com) |
2 | * All rights reserved. |
3 | * |
4 | * This package is an SSL implementation written |
5 | * by Eric Young (eay@cryptsoft.com). |
6 | * The implementation was written so as to conform with Netscapes SSL. |
7 | * |
8 | * This library is free for commercial and non-commercial use as long as |
9 | * the following conditions are aheared to. The following conditions |
10 | * apply to all code found in this distribution, be it the RC4, RSA, |
11 | * lhash, DES, etc., code; not just the SSL code. The SSL documentation |
12 | * included with this distribution is covered by the same copyright terms |
13 | * except that the holder is Tim Hudson (tjh@cryptsoft.com). |
14 | * |
15 | * Copyright remains Eric Young's, and as such any Copyright notices in |
16 | * the code are not to be removed. |
17 | * If this package is used in a product, Eric Young should be given attribution |
18 | * as the author of the parts of the library used. |
19 | * This can be in the form of a textual message at program startup or |
20 | * in documentation (online or textual) provided with the package. |
21 | * |
22 | * Redistribution and use in source and binary forms, with or without |
23 | * modification, are permitted provided that the following conditions |
24 | * are met: |
25 | * 1. Redistributions of source code must retain the copyright |
26 | * notice, this list of conditions and the following disclaimer. |
27 | * 2. Redistributions in binary form must reproduce the above copyright |
28 | * notice, this list of conditions and the following disclaimer in the |
29 | * documentation and/or other materials provided with the distribution. |
30 | * 3. All advertising materials mentioning features or use of this software |
31 | * must display the following acknowledgement: |
32 | * "This product includes cryptographic software written by |
33 | * Eric Young (eay@cryptsoft.com)" |
34 | * The word 'cryptographic' can be left out if the rouines from the library |
35 | * being used are not cryptographic related :-). |
36 | * 4. If you include any Windows specific code (or a derivative thereof) from |
37 | * the apps directory (application code) you must include an acknowledgement: |
38 | * "This product includes software written by Tim Hudson (tjh@cryptsoft.com)" |
39 | * |
40 | * THIS SOFTWARE IS PROVIDED BY ERIC YOUNG ``AS IS'' AND |
41 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
42 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
43 | * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE |
44 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
45 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
46 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
47 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
48 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
49 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
50 | * SUCH DAMAGE. |
51 | * |
52 | * The licence and distribution terms for any publically available version or |
53 | * derivative of this code cannot be changed. i.e. this code cannot simply be |
54 | * copied and put under another distribution licence |
55 | * [including the GNU Public Licence.] */ |
56 | |
57 | #include <openssl/cpu.h> |
58 | |
59 | |
60 | #if !defined(OPENSSL_NO_ASM) && (defined(OPENSSL_X86) || defined(OPENSSL_X86_64)) |
61 | |
62 | #include <inttypes.h> |
63 | #include <stdio.h> |
64 | #include <stdlib.h> |
65 | #include <string.h> |
66 | |
67 | #if defined(_MSC_VER) |
68 | OPENSSL_MSVC_PRAGMA(warning(push, 3)) |
69 | #include <immintrin.h> |
70 | #include <intrin.h> |
71 | OPENSSL_MSVC_PRAGMA(warning(pop)) |
72 | #endif |
73 | |
74 | #include "internal.h" |
75 | |
76 | |
77 | // OPENSSL_cpuid runs the cpuid instruction. |leaf| is passed in as EAX and ECX |
78 | // is set to zero. It writes EAX, EBX, ECX, and EDX to |*out_eax| through |
79 | // |*out_edx|. |
80 | static void OPENSSL_cpuid(uint32_t *out_eax, uint32_t *out_ebx, |
81 | uint32_t *out_ecx, uint32_t *out_edx, uint32_t leaf) { |
82 | #if defined(_MSC_VER) |
83 | int tmp[4]; |
84 | __cpuid(tmp, (int)leaf); |
85 | *out_eax = (uint32_t)tmp[0]; |
86 | *out_ebx = (uint32_t)tmp[1]; |
87 | *out_ecx = (uint32_t)tmp[2]; |
88 | *out_edx = (uint32_t)tmp[3]; |
89 | #elif defined(__pic__) && defined(OPENSSL_32_BIT) |
90 | // Inline assembly may not clobber the PIC register. For 32-bit, this is EBX. |
91 | // See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=47602. |
92 | __asm__ volatile ( |
93 | "xor %%ecx, %%ecx\n" |
94 | "mov %%ebx, %%edi\n" |
95 | "cpuid\n" |
96 | "xchg %%edi, %%ebx\n" |
97 | : "=a" (*out_eax), "=D" (*out_ebx), "=c" (*out_ecx), "=d" (*out_edx) |
98 | : "a" (leaf) |
99 | ); |
100 | #else |
101 | __asm__ volatile ( |
102 | "xor %%ecx, %%ecx\n" |
103 | "cpuid\n" |
104 | : "=a" (*out_eax), "=b" (*out_ebx), "=c" (*out_ecx), "=d" (*out_edx) |
105 | : "a" (leaf) |
106 | ); |
107 | #endif |
108 | } |
109 | |
110 | // OPENSSL_xgetbv returns the value of an Intel Extended Control Register (XCR). |
111 | // Currently only XCR0 is defined by Intel so |xcr| should always be zero. |
112 | static uint64_t OPENSSL_xgetbv(uint32_t xcr) { |
113 | #if defined(_MSC_VER) |
114 | return (uint64_t)_xgetbv(xcr); |
115 | #else |
116 | uint32_t eax, edx; |
117 | __asm__ volatile ("xgetbv" : "=a" (eax), "=d" (edx) : "c" (xcr)); |
118 | return (((uint64_t)edx) << 32) | eax; |
119 | #endif |
120 | } |
121 | |
122 | // handle_cpu_env applies the value from |in| to the CPUID values in |out[0]| |
123 | // and |out[1]|. See the comment in |OPENSSL_cpuid_setup| about this. |
124 | static void handle_cpu_env(uint32_t *out, const char *in) { |
125 | const int invert = in[0] == '~'; |
126 | uint64_t v; |
127 | |
128 | if (!sscanf(in + invert, "%" PRIu64, &v)) { |
129 | return; |
130 | } |
131 | |
132 | if (invert) { |
133 | out[0] &= ~v; |
134 | out[1] &= ~(v >> 32); |
135 | } else { |
136 | out[0] = v; |
137 | out[1] = v >> 32; |
138 | } |
139 | } |
140 | |
141 | void OPENSSL_cpuid_setup(void) { |
142 | // Determine the vendor and maximum input value. |
143 | uint32_t eax, ebx, ecx, edx; |
144 | OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 0); |
145 | |
146 | uint32_t num_ids = eax; |
147 | |
148 | int is_intel = ebx == 0x756e6547 /* Genu */ && |
149 | edx == 0x49656e69 /* ineI */ && |
150 | ecx == 0x6c65746e /* ntel */; |
151 | int is_amd = ebx == 0x68747541 /* Auth */ && |
152 | edx == 0x69746e65 /* enti */ && |
153 | ecx == 0x444d4163 /* cAMD */; |
154 | |
155 | uint32_t extended_features[2] = {0}; |
156 | if (num_ids >= 7) { |
157 | OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 7); |
158 | extended_features[0] = ebx; |
159 | extended_features[1] = ecx; |
160 | } |
161 | |
162 | OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 1); |
163 | |
164 | if (is_amd) { |
165 | // See https://www.amd.com/system/files/TechDocs/25481.pdf, page 10. |
166 | const uint32_t base_family = (eax >> 8) & 15; |
167 | |
168 | uint32_t family = base_family; |
169 | if (base_family == 0xf) { |
170 | const uint32_t ext_family = (eax >> 20) & 255; |
171 | family += ext_family; |
172 | } |
173 | |
174 | if (family < 0x17) { |
175 | // Disable RDRAND on AMD families before 0x17 (Zen) due to reported |
176 | // failures after suspend. |
177 | // https://bugzilla.redhat.com/show_bug.cgi?id=1150286 |
178 | ecx &= ~(1u << 30); |
179 | } |
180 | } |
181 | |
182 | // Force the hyper-threading bit so that the more conservative path is always |
183 | // chosen. |
184 | edx |= 1u << 28; |
185 | |
186 | // Reserved bit #20 was historically repurposed to control the in-memory |
187 | // representation of RC4 state. Always set it to zero. |
188 | edx &= ~(1u << 20); |
189 | |
190 | // Reserved bit #30 is repurposed to signal an Intel CPU. |
191 | if (is_intel) { |
192 | edx |= (1u << 30); |
193 | |
194 | // Clear the XSAVE bit on Knights Landing to mimic Silvermont. This enables |
195 | // some Silvermont-specific codepaths which perform better. See OpenSSL |
196 | // commit 64d92d74985ebb3d0be58a9718f9e080a14a8e7f. |
197 | if ((eax & 0x0fff0ff0) == 0x00050670 /* Knights Landing */ || |
198 | (eax & 0x0fff0ff0) == 0x00080650 /* Knights Mill (per SDE) */) { |
199 | ecx &= ~(1u << 26); |
200 | } |
201 | } else { |
202 | edx &= ~(1u << 30); |
203 | } |
204 | |
205 | // The SDBG bit is repurposed to denote AMD XOP support. Don't ever use AMD |
206 | // XOP code paths. |
207 | ecx &= ~(1u << 11); |
208 | |
209 | uint64_t xcr0 = 0; |
210 | if (ecx & (1u << 27)) { |
211 | // XCR0 may only be queried if the OSXSAVE bit is set. |
212 | xcr0 = OPENSSL_xgetbv(0); |
213 | } |
214 | // See Intel manual, volume 1, section 14.3. |
215 | if ((xcr0 & 6) != 6) { |
216 | // YMM registers cannot be used. |
217 | ecx &= ~(1u << 28); // AVX |
218 | ecx &= ~(1u << 12); // FMA |
219 | ecx &= ~(1u << 11); // AMD XOP |
220 | // Clear AVX2 and AVX512* bits. |
221 | // |
222 | // TODO(davidben): Should bits 17 and 26-28 also be cleared? Upstream |
223 | // doesn't clear those. |
224 | extended_features[0] &= |
225 | ~((1u << 5) | (1u << 16) | (1u << 21) | (1u << 30) | (1u << 31)); |
226 | } |
227 | // See Intel manual, volume 1, section 15.2. |
228 | if ((xcr0 & 0xe6) != 0xe6) { |
229 | // Clear AVX512F. Note we don't touch other AVX512 extensions because they |
230 | // can be used with YMM. |
231 | extended_features[0] &= ~(1u << 16); |
232 | } |
233 | |
234 | // Disable ADX instructions on Knights Landing. See OpenSSL commit |
235 | // 64d92d74985ebb3d0be58a9718f9e080a14a8e7f. |
236 | if ((ecx & (1u << 26)) == 0) { |
237 | extended_features[0] &= ~(1u << 19); |
238 | } |
239 | |
240 | OPENSSL_ia32cap_P[0] = edx; |
241 | OPENSSL_ia32cap_P[1] = ecx; |
242 | OPENSSL_ia32cap_P[2] = extended_features[0]; |
243 | OPENSSL_ia32cap_P[3] = extended_features[1]; |
244 | |
245 | const char *env1, *env2; |
246 | env1 = getenv("OPENSSL_ia32cap" ); |
247 | if (env1 == NULL) { |
248 | return; |
249 | } |
250 | |
251 | // OPENSSL_ia32cap can contain zero, one or two values, separated with a ':'. |
252 | // Each value is a 64-bit, unsigned value which may start with "0x" to |
253 | // indicate a hex value. Prior to the 64-bit value, a '~' may be given. |
254 | // |
255 | // If '~' isn't present, then the value is taken as the result of the CPUID. |
256 | // Otherwise the value is inverted and ANDed with the probed CPUID result. |
257 | // |
258 | // The first value determines OPENSSL_ia32cap_P[0] and [1]. The second [2] |
259 | // and [3]. |
260 | |
261 | handle_cpu_env(&OPENSSL_ia32cap_P[0], env1); |
262 | env2 = strchr(env1, ':'); |
263 | if (env2 != NULL) { |
264 | handle_cpu_env(&OPENSSL_ia32cap_P[2], env2 + 1); |
265 | } |
266 | } |
267 | |
268 | #endif // !OPENSSL_NO_ASM && (OPENSSL_X86 || OPENSSL_X86_64) |
269 | |