xref: /aosp_15_r20/external/boringssl/src/crypto/cpu_intel.c (revision 8fb009dc861624b67b6cdb62ea21f0f22d0c584b)
1 /* Copyright (C) 1995-1998 Eric Young ([email protected])
2  * All rights reserved.
3  *
4  * This package is an SSL implementation written
5  * by Eric Young ([email protected]).
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 ([email protected]).
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 ([email protected])"
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 ([email protected])"
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/base.h>
58 
59 #if !defined(OPENSSL_NO_ASM) && (defined(OPENSSL_X86) || defined(OPENSSL_X86_64))
60 
61 #include <inttypes.h>
62 #include <stdio.h>
63 #include <stdlib.h>
64 #include <string.h>
65 
66 #if defined(_MSC_VER)
67 OPENSSL_MSVC_PRAGMA(warning(push, 3))
68 #include <immintrin.h>
69 #include <intrin.h>
OPENSSL_MSVC_PRAGMA(warning (pop))70 OPENSSL_MSVC_PRAGMA(warning(pop))
71 #endif
72 
73 #include "internal.h"
74 
75 
76 // OPENSSL_cpuid runs the cpuid instruction. |leaf| is passed in as EAX and ECX
77 // is set to zero. It writes EAX, EBX, ECX, and EDX to |*out_eax| through
78 // |*out_edx|.
79 static void OPENSSL_cpuid(uint32_t *out_eax, uint32_t *out_ebx,
80                           uint32_t *out_ecx, uint32_t *out_edx, uint32_t leaf) {
81 #if defined(_MSC_VER)
82   int tmp[4];
83   __cpuid(tmp, (int)leaf);
84   *out_eax = (uint32_t)tmp[0];
85   *out_ebx = (uint32_t)tmp[1];
86   *out_ecx = (uint32_t)tmp[2];
87   *out_edx = (uint32_t)tmp[3];
88 #elif defined(__pic__) && defined(OPENSSL_32_BIT)
89   // Inline assembly may not clobber the PIC register. For 32-bit, this is EBX.
90   // See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=47602.
91   __asm__ volatile (
92     "xor %%ecx, %%ecx\n"
93     "mov %%ebx, %%edi\n"
94     "cpuid\n"
95     "xchg %%edi, %%ebx\n"
96     : "=a"(*out_eax), "=D"(*out_ebx), "=c"(*out_ecx), "=d"(*out_edx)
97     : "a"(leaf)
98   );
99 #else
100   __asm__ volatile (
101     "xor %%ecx, %%ecx\n"
102     "cpuid\n"
103     : "=a"(*out_eax), "=b"(*out_ebx), "=c"(*out_ecx), "=d"(*out_edx)
104     : "a"(leaf)
105   );
106 #endif
107 }
108 
109 // OPENSSL_xgetbv returns the value of an Intel Extended Control Register (XCR).
110 // Currently only XCR0 is defined by Intel so |xcr| should always be zero.
OPENSSL_xgetbv(uint32_t xcr)111 static uint64_t OPENSSL_xgetbv(uint32_t xcr) {
112 #if defined(_MSC_VER)
113   return (uint64_t)_xgetbv(xcr);
114 #else
115   uint32_t eax, edx;
116   __asm__ volatile ("xgetbv" : "=a"(eax), "=d"(edx) : "c"(xcr));
117   return (((uint64_t)edx) << 32) | eax;
118 #endif
119 }
120 
121 // handle_cpu_env applies the value from |in| to the CPUID values in |out[0]|
122 // and |out[1]|. See the comment in |OPENSSL_cpuid_setup| about this.
handle_cpu_env(uint32_t * out,const char * in)123 static void handle_cpu_env(uint32_t *out, const char *in) {
124   const int invert = in[0] == '~';
125   const int or = in[0] == '|';
126   const int skip_first_byte = invert || or;
127   const int hex = in[skip_first_byte] == '0' && in[skip_first_byte+1] == 'x';
128 
129   int sscanf_result;
130   uint64_t v;
131   if (hex) {
132     sscanf_result = sscanf(in + invert + 2, "%" PRIx64, &v);
133   } else {
134     sscanf_result = sscanf(in + invert, "%" PRIu64, &v);
135   }
136 
137   if (!sscanf_result) {
138     return;
139   }
140 
141   if (invert) {
142     out[0] &= ~v;
143     out[1] &= ~(v >> 32);
144   } else if (or) {
145     out[0] |= v;
146     out[1] |= (v >> 32);
147   } else {
148     out[0] = v;
149     out[1] = v >> 32;
150   }
151 }
152 
OPENSSL_cpuid_setup(void)153 void OPENSSL_cpuid_setup(void) {
154   // Determine the vendor and maximum input value.
155   uint32_t eax, ebx, ecx, edx;
156   OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 0);
157 
158   uint32_t num_ids = eax;
159 
160   int is_intel = ebx == 0x756e6547 /* Genu */ &&
161                  edx == 0x49656e69 /* ineI */ &&
162                  ecx == 0x6c65746e /* ntel */;
163   int is_amd = ebx == 0x68747541 /* Auth */ &&
164                edx == 0x69746e65 /* enti */ &&
165                ecx == 0x444d4163 /* cAMD */;
166 
167   uint32_t extended_features[2] = {0};
168   if (num_ids >= 7) {
169     OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 7);
170     extended_features[0] = ebx;
171     extended_features[1] = ecx;
172   }
173 
174   OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 1);
175 
176   if (is_amd) {
177     // See https://www.amd.com/system/files/TechDocs/25481.pdf, page 10.
178     const uint32_t base_family = (eax >> 8) & 15;
179     const uint32_t base_model = (eax >> 4) & 15;
180 
181     uint32_t family = base_family;
182     uint32_t model = base_model;
183     if (base_family == 0xf) {
184       const uint32_t ext_family = (eax >> 20) & 255;
185       family += ext_family;
186       const uint32_t ext_model = (eax >> 16) & 15;
187       model |= ext_model << 4;
188     }
189 
190     if (family < 0x17 || (family == 0x17 && 0x70 <= model && model <= 0x7f)) {
191       // Disable RDRAND on AMD families before 0x17 (Zen) due to reported
192       // failures after suspend.
193       // https://bugzilla.redhat.com/show_bug.cgi?id=1150286
194       // Also disable for family 0x17, models 0x70–0x7f, due to possible RDRAND
195       // failures there too.
196       ecx &= ~(1u << 30);
197     }
198   }
199 
200   // Force the hyper-threading bit so that the more conservative path is always
201   // chosen.
202   edx |= 1u << 28;
203 
204   // Reserved bit #20 was historically repurposed to control the in-memory
205   // representation of RC4 state. Always set it to zero.
206   edx &= ~(1u << 20);
207 
208   // Reserved bit #30 is repurposed to signal an Intel CPU.
209   if (is_intel) {
210     edx |= (1u << 30);
211   } else {
212     edx &= ~(1u << 30);
213   }
214 
215   // The SDBG bit is repurposed to denote AMD XOP support. Don't ever use AMD
216   // XOP code paths.
217   ecx &= ~(1u << 11);
218 
219   uint64_t xcr0 = 0;
220   if (ecx & (1u << 27)) {
221     // XCR0 may only be queried if the OSXSAVE bit is set.
222     xcr0 = OPENSSL_xgetbv(0);
223   }
224   // See Intel manual, volume 1, section 14.3.
225   if ((xcr0 & 6) != 6) {
226     // YMM registers cannot be used.
227     ecx &= ~(1u << 28);  // AVX
228     ecx &= ~(1u << 12);  // FMA
229     ecx &= ~(1u << 11);  // AMD XOP
230     // Clear AVX2 and AVX512* bits.
231     //
232     // TODO(davidben): Should bits 17 and 26-28 also be cleared? Upstream
233     // doesn't clear those. See the comments in
234     // |CRYPTO_hardware_supports_XSAVE|.
235     extended_features[0] &=
236         ~((1u << 5) | (1u << 16) | (1u << 21) | (1u << 30) | (1u << 31));
237   }
238   // See Intel manual, volume 1, section 15.2.
239   if ((xcr0 & 0xe6) != 0xe6) {
240     // Clear AVX512F. Note we don't touch other AVX512 extensions because they
241     // can be used with YMM.
242     extended_features[0] &= ~(1u << 16);
243   }
244 
245   OPENSSL_ia32cap_P[0] = edx;
246   OPENSSL_ia32cap_P[1] = ecx;
247   OPENSSL_ia32cap_P[2] = extended_features[0];
248   OPENSSL_ia32cap_P[3] = extended_features[1];
249 
250   const char *env1, *env2;
251   env1 = getenv("OPENSSL_ia32cap");
252   if (env1 == NULL) {
253     return;
254   }
255 
256   // OPENSSL_ia32cap can contain zero, one or two values, separated with a ':'.
257   // Each value is a 64-bit, unsigned value which may start with "0x" to
258   // indicate a hex value. Prior to the 64-bit value, a '~' or '|' may be given.
259   //
260   // If the '~' prefix is present:
261   //   the value is inverted and ANDed with the probed CPUID result
262   // If the '|' prefix is present:
263   //   the value is ORed with the probed CPUID result
264   // Otherwise:
265   //   the value is taken as the result of the CPUID
266   //
267   // The first value determines OPENSSL_ia32cap_P[0] and [1]. The second [2]
268   // and [3].
269 
270   handle_cpu_env(&OPENSSL_ia32cap_P[0], env1);
271   env2 = strchr(env1, ':');
272   if (env2 != NULL) {
273     handle_cpu_env(&OPENSSL_ia32cap_P[2], env2 + 1);
274   }
275 }
276 
277 #endif  // !OPENSSL_NO_ASM && (OPENSSL_X86 || OPENSSL_X86_64)
278