• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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>
OPENSSL_MSVC_PRAGMA(warning (pop))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.
OPENSSL_xgetbv(uint32_t xcr)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.
handle_cpu_env(uint32_t * out,const char * in)124 static void handle_cpu_env(uint32_t *out, const char *in) {
125   const int invert = in[0] == '~';
126   const int or = in[0] == '|';
127   const int skip_first_byte = invert || or;
128   const int hex = in[skip_first_byte] == '0' && in[skip_first_byte+1] == 'x';
129 
130   int sscanf_result;
131   uint64_t v;
132   if (hex) {
133     sscanf_result = sscanf(in + invert + 2, "%" PRIx64, &v);
134   } else {
135     sscanf_result = sscanf(in + invert, "%" PRIu64, &v);
136   }
137 
138   if (!sscanf_result) {
139     return;
140   }
141 
142   if (invert) {
143     out[0] &= ~v;
144     out[1] &= ~(v >> 32);
145   } else if (or) {
146     out[0] |= v;
147     out[1] |= (v >> 32);
148   } else {
149     out[0] = v;
150     out[1] = v >> 32;
151   }
152 }
153 
OPENSSL_cpuid_setup(void)154 void OPENSSL_cpuid_setup(void) {
155   // Determine the vendor and maximum input value.
156   uint32_t eax, ebx, ecx, edx;
157   OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 0);
158 
159   uint32_t num_ids = eax;
160 
161   int is_intel = ebx == 0x756e6547 /* Genu */ &&
162                  edx == 0x49656e69 /* ineI */ &&
163                  ecx == 0x6c65746e /* ntel */;
164   int is_amd = ebx == 0x68747541 /* Auth */ &&
165                edx == 0x69746e65 /* enti */ &&
166                ecx == 0x444d4163 /* cAMD */;
167 
168   uint32_t extended_features[2] = {0};
169   if (num_ids >= 7) {
170     OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 7);
171     extended_features[0] = ebx;
172     extended_features[1] = ecx;
173   }
174 
175   OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 1);
176 
177   if (is_amd) {
178     // See https://www.amd.com/system/files/TechDocs/25481.pdf, page 10.
179     const uint32_t base_family = (eax >> 8) & 15;
180     const uint32_t base_model = (eax >> 4) & 15;
181 
182     uint32_t family = base_family;
183     uint32_t model = base_model;
184     if (base_family == 0xf) {
185       const uint32_t ext_family = (eax >> 20) & 255;
186       family += ext_family;
187       const uint32_t ext_model = (eax >> 16) & 15;
188       model |= ext_model << 4;
189     }
190 
191     if (family < 0x17 || (family == 0x17 && 0x70 <= model && model <= 0x7f)) {
192       // Disable RDRAND on AMD families before 0x17 (Zen) due to reported
193       // failures after suspend.
194       // https://bugzilla.redhat.com/show_bug.cgi?id=1150286
195       // Also disable for family 0x17, models 0x70–0x7f, due to possible RDRAND
196       // failures there too.
197       ecx &= ~(1u << 30);
198     }
199   }
200 
201   // Force the hyper-threading bit so that the more conservative path is always
202   // chosen.
203   edx |= 1u << 28;
204 
205   // Reserved bit #20 was historically repurposed to control the in-memory
206   // representation of RC4 state. Always set it to zero.
207   edx &= ~(1u << 20);
208 
209   // Reserved bit #30 is repurposed to signal an Intel CPU.
210   if (is_intel) {
211     edx |= (1u << 30);
212 
213     // Clear the XSAVE bit on Knights Landing to mimic Silvermont. This enables
214     // some Silvermont-specific codepaths which perform better. See OpenSSL
215     // commit 64d92d74985ebb3d0be58a9718f9e080a14a8e7f.
216     if ((eax & 0x0fff0ff0) == 0x00050670 /* Knights Landing */ ||
217         (eax & 0x0fff0ff0) == 0x00080650 /* Knights Mill (per SDE) */) {
218       ecx &= ~(1u << 26);
219     }
220   } else {
221     edx &= ~(1u << 30);
222   }
223 
224   // The SDBG bit is repurposed to denote AMD XOP support. Don't ever use AMD
225   // XOP code paths.
226   ecx &= ~(1u << 11);
227 
228   uint64_t xcr0 = 0;
229   if (ecx & (1u << 27)) {
230     // XCR0 may only be queried if the OSXSAVE bit is set.
231     xcr0 = OPENSSL_xgetbv(0);
232   }
233   // See Intel manual, volume 1, section 14.3.
234   if ((xcr0 & 6) != 6) {
235     // YMM registers cannot be used.
236     ecx &= ~(1u << 28);  // AVX
237     ecx &= ~(1u << 12);  // FMA
238     ecx &= ~(1u << 11);  // AMD XOP
239     // Clear AVX2 and AVX512* bits.
240     //
241     // TODO(davidben): Should bits 17 and 26-28 also be cleared? Upstream
242     // doesn't clear those.
243     extended_features[0] &=
244         ~((1u << 5) | (1u << 16) | (1u << 21) | (1u << 30) | (1u << 31));
245   }
246   // See Intel manual, volume 1, section 15.2.
247   if ((xcr0 & 0xe6) != 0xe6) {
248     // Clear AVX512F. Note we don't touch other AVX512 extensions because they
249     // can be used with YMM.
250     extended_features[0] &= ~(1u << 16);
251   }
252 
253   // Disable ADX instructions on Knights Landing. See OpenSSL commit
254   // 64d92d74985ebb3d0be58a9718f9e080a14a8e7f.
255   if ((ecx & (1u << 26)) == 0) {
256     extended_features[0] &= ~(1u << 19);
257   }
258 
259   OPENSSL_ia32cap_P[0] = edx;
260   OPENSSL_ia32cap_P[1] = ecx;
261   OPENSSL_ia32cap_P[2] = extended_features[0];
262   OPENSSL_ia32cap_P[3] = extended_features[1];
263 
264   const char *env1, *env2;
265   env1 = getenv("OPENSSL_ia32cap");
266   if (env1 == NULL) {
267     return;
268   }
269 
270   // OPENSSL_ia32cap can contain zero, one or two values, separated with a ':'.
271   // Each value is a 64-bit, unsigned value which may start with "0x" to
272   // indicate a hex value. Prior to the 64-bit value, a '~' or '|' may be given.
273   //
274   // If the '~' prefix is present:
275   //   the value is inverted and ANDed with the probed CPUID result
276   // If the '|' prefix is present:
277   //   the value is ORed with the probed CPUID result
278   // Otherwise:
279   //   the value is taken as the result of the CPUID
280   //
281   // The first value determines OPENSSL_ia32cap_P[0] and [1]. The second [2]
282   // and [3].
283 
284   handle_cpu_env(&OPENSSL_ia32cap_P[0], env1);
285   env2 = strchr(env1, ':');
286   if (env2 != NULL) {
287     handle_cpu_env(&OPENSSL_ia32cap_P[2], env2 + 1);
288   }
289 }
290 
291 #endif  // !OPENSSL_NO_ASM && (OPENSSL_X86 || OPENSSL_X86_64)
292