• 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   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 
OPENSSL_cpuid_setup(void)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 
152   uint32_t extended_features[2] = {0};
153   if (num_ids >= 7) {
154     OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 7);
155     extended_features[0] = ebx;
156     extended_features[1] = ecx;
157   }
158 
159   OPENSSL_cpuid(&eax, &ebx, &ecx, &edx, 1);
160 
161   // Force the hyper-threading bit so that the more conservative path is always
162   // chosen.
163   edx |= 1u << 28;
164 
165   // Reserved bit #20 was historically repurposed to control the in-memory
166   // representation of RC4 state. Always set it to zero.
167   edx &= ~(1u << 20);
168 
169   // Reserved bit #30 is repurposed to signal an Intel CPU.
170   if (is_intel) {
171     edx |= (1u << 30);
172 
173     // Clear the XSAVE bit on Knights Landing to mimic Silvermont. This enables
174     // some Silvermont-specific codepaths which perform better. See OpenSSL
175     // commit 64d92d74985ebb3d0be58a9718f9e080a14a8e7f.
176     if ((eax & 0x0fff0ff0) == 0x00050670 /* Knights Landing */ ||
177         (eax & 0x0fff0ff0) == 0x00080650 /* Knights Mill (per SDE) */) {
178       ecx &= ~(1u << 26);
179     }
180   } else {
181     edx &= ~(1u << 30);
182   }
183 
184   // The SDBG bit is repurposed to denote AMD XOP support. Don't ever use AMD
185   // XOP code paths.
186   ecx &= ~(1u << 11);
187 
188   uint64_t xcr0 = 0;
189   if (ecx & (1u << 27)) {
190     // XCR0 may only be queried if the OSXSAVE bit is set.
191     xcr0 = OPENSSL_xgetbv(0);
192   }
193   // See Intel manual, volume 1, section 14.3.
194   if ((xcr0 & 6) != 6) {
195     // YMM registers cannot be used.
196     ecx &= ~(1u << 28);  // AVX
197     ecx &= ~(1u << 12);  // FMA
198     ecx &= ~(1u << 11);  // AMD XOP
199     // Clear AVX2 and AVX512* bits.
200     //
201     // TODO(davidben): Should bits 17 and 26-28 also be cleared? Upstream
202     // doesn't clear those.
203     extended_features[0] &=
204         ~((1u << 5) | (1u << 16) | (1u << 21) | (1u << 30) | (1u << 31));
205   }
206   // See Intel manual, volume 1, section 15.2.
207   if ((xcr0 & 0xe6) != 0xe6) {
208     // Clear AVX512F. Note we don't touch other AVX512 extensions because they
209     // can be used with YMM.
210     extended_features[0] &= ~(1u << 16);
211   }
212 
213   // Disable ADX instructions on Knights Landing. See OpenSSL commit
214   // 64d92d74985ebb3d0be58a9718f9e080a14a8e7f.
215   if ((ecx & (1u << 26)) == 0) {
216     extended_features[0] &= ~(1u << 19);
217   }
218 
219   OPENSSL_ia32cap_P[0] = edx;
220   OPENSSL_ia32cap_P[1] = ecx;
221   OPENSSL_ia32cap_P[2] = extended_features[0];
222   OPENSSL_ia32cap_P[3] = extended_features[1];
223 
224   const char *env1, *env2;
225   env1 = getenv("OPENSSL_ia32cap");
226   if (env1 == NULL) {
227     return;
228   }
229 
230   // OPENSSL_ia32cap can contain zero, one or two values, separated with a ':'.
231   // Each value is a 64-bit, unsigned value which may start with "0x" to
232   // indicate a hex value. Prior to the 64-bit value, a '~' may be given.
233   //
234   // If '~' isn't present, then the value is taken as the result of the CPUID.
235   // Otherwise the value is inverted and ANDed with the probed CPUID result.
236   //
237   // The first value determines OPENSSL_ia32cap_P[0] and [1]. The second [2]
238   // and [3].
239 
240   handle_cpu_env(&OPENSSL_ia32cap_P[0], env1);
241   env2 = strchr(env1, ':');
242   if (env2 != NULL) {
243     handle_cpu_env(&OPENSSL_ia32cap_P[2], env2 + 1);
244   }
245 }
246 
247 #endif  // !OPENSSL_NO_ASM && (OPENSSL_X86 || OPENSSL_X86_64)
248