xref: /third_party/mbedtls/library/aesni.c (revision a8e1175b)
1/*
2 *  AES-NI support functions
3 *
4 *  Copyright The Mbed TLS Contributors
5 *  SPDX-License-Identifier: Apache-2.0 OR GPL-2.0-or-later
6 */
7
8/*
9 * [AES-WP] https://www.intel.com/content/www/us/en/developer/articles/tool/intel-advanced-encryption-standard-aes-instructions-set.html
10 * [CLMUL-WP] https://www.intel.com/content/www/us/en/develop/download/intel-carry-less-multiplication-instruction-and-its-usage-for-computing-the-gcm-mode.html
11 */
12
13#include "common.h"
14
15#if defined(MBEDTLS_AESNI_C)
16
17#include "aesni.h"
18
19#include <string.h>
20
21#if defined(MBEDTLS_AESNI_HAVE_CODE)
22
23#if MBEDTLS_AESNI_HAVE_CODE == 2
24#if defined(__GNUC__)
25#include <cpuid.h>
26#elif defined(_MSC_VER)
27#include <intrin.h>
28#else
29#error "`__cpuid` required by MBEDTLS_AESNI_C is not supported by the compiler"
30#endif
31#include <immintrin.h>
32#endif
33
34#if defined(MBEDTLS_ARCH_IS_X86)
35#if defined(MBEDTLS_COMPILER_IS_GCC)
36#pragma GCC push_options
37#pragma GCC target ("pclmul,sse2,aes")
38#define MBEDTLS_POP_TARGET_PRAGMA
39#elif defined(__clang__) && (__clang_major__ >= 5)
40#pragma clang attribute push (__attribute__((target("pclmul,sse2,aes"))), apply_to=function)
41#define MBEDTLS_POP_TARGET_PRAGMA
42#endif
43#endif
44
45#if !defined(MBEDTLS_AES_USE_HARDWARE_ONLY)
46/*
47 * AES-NI support detection routine
48 */
49int mbedtls_aesni_has_support(unsigned int what)
50{
51    static int done = 0;
52    static unsigned int c = 0;
53
54    if (!done) {
55#if MBEDTLS_AESNI_HAVE_CODE == 2
56        static int info[4] = { 0, 0, 0, 0 };
57#if defined(_MSC_VER)
58        __cpuid(info, 1);
59#else
60        __cpuid(1, info[0], info[1], info[2], info[3]);
61#endif
62        c = info[2];
63#else /* AESNI using asm */
64        asm ("movl  $1, %%eax   \n\t"
65             "cpuid             \n\t"
66             : "=c" (c)
67             :
68             : "eax", "ebx", "edx");
69#endif /* MBEDTLS_AESNI_HAVE_CODE */
70        done = 1;
71    }
72
73    return (c & what) != 0;
74}
75#endif /* !MBEDTLS_AES_USE_HARDWARE_ONLY */
76
77#if MBEDTLS_AESNI_HAVE_CODE == 2
78
79/*
80 * AES-NI AES-ECB block en(de)cryption
81 */
82int mbedtls_aesni_crypt_ecb(mbedtls_aes_context *ctx,
83                            int mode,
84                            const unsigned char input[16],
85                            unsigned char output[16])
86{
87    const __m128i *rk = (const __m128i *) (ctx->buf + ctx->rk_offset);
88    unsigned nr = ctx->nr; // Number of remaining rounds
89
90    // Load round key 0
91    __m128i state;
92    memcpy(&state, input, 16);
93    state = _mm_xor_si128(state, rk[0]);  // state ^= *rk;
94    ++rk;
95    --nr;
96
97#if !defined(MBEDTLS_BLOCK_CIPHER_NO_DECRYPT)
98    if (mode == MBEDTLS_AES_DECRYPT) {
99        while (nr != 0) {
100            state = _mm_aesdec_si128(state, *rk);
101            ++rk;
102            --nr;
103        }
104        state = _mm_aesdeclast_si128(state, *rk);
105    } else
106#else
107    (void) mode;
108#endif
109    {
110        while (nr != 0) {
111            state = _mm_aesenc_si128(state, *rk);
112            ++rk;
113            --nr;
114        }
115        state = _mm_aesenclast_si128(state, *rk);
116    }
117
118    memcpy(output, &state, 16);
119    return 0;
120}
121
122/*
123 * GCM multiplication: c = a times b in GF(2^128)
124 * Based on [CLMUL-WP] algorithms 1 (with equation 27) and 5.
125 */
126
127static void gcm_clmul(const __m128i aa, const __m128i bb,
128                      __m128i *cc, __m128i *dd)
129{
130    /*
131     * Caryless multiplication dd:cc = aa * bb
132     * using [CLMUL-WP] algorithm 1 (p. 12).
133     */
134    *cc = _mm_clmulepi64_si128(aa, bb, 0x00); // a0*b0 = c1:c0
135    *dd = _mm_clmulepi64_si128(aa, bb, 0x11); // a1*b1 = d1:d0
136    __m128i ee = _mm_clmulepi64_si128(aa, bb, 0x10); // a0*b1 = e1:e0
137    __m128i ff = _mm_clmulepi64_si128(aa, bb, 0x01); // a1*b0 = f1:f0
138    ff = _mm_xor_si128(ff, ee);                      // e1+f1:e0+f0
139    ee = ff;                                         // e1+f1:e0+f0
140    ff = _mm_srli_si128(ff, 8);                      // 0:e1+f1
141    ee = _mm_slli_si128(ee, 8);                      // e0+f0:0
142    *dd = _mm_xor_si128(*dd, ff);                    // d1:d0+e1+f1
143    *cc = _mm_xor_si128(*cc, ee);                    // c1+e0+f0:c0
144}
145
146static void gcm_shift(__m128i *cc, __m128i *dd)
147{
148    /* [CMUCL-WP] Algorithm 5 Step 1: shift cc:dd one bit to the left,
149     * taking advantage of [CLMUL-WP] eq 27 (p. 18). */
150    //                                        // *cc = r1:r0
151    //                                        // *dd = r3:r2
152    __m128i cc_lo = _mm_slli_epi64(*cc, 1);   // r1<<1:r0<<1
153    __m128i dd_lo = _mm_slli_epi64(*dd, 1);   // r3<<1:r2<<1
154    __m128i cc_hi = _mm_srli_epi64(*cc, 63);  // r1>>63:r0>>63
155    __m128i dd_hi = _mm_srli_epi64(*dd, 63);  // r3>>63:r2>>63
156    __m128i xmm5 = _mm_srli_si128(cc_hi, 8);  // 0:r1>>63
157    cc_hi = _mm_slli_si128(cc_hi, 8);         // r0>>63:0
158    dd_hi = _mm_slli_si128(dd_hi, 8);         // 0:r1>>63
159
160    *cc = _mm_or_si128(cc_lo, cc_hi);         // r1<<1|r0>>63:r0<<1
161    *dd = _mm_or_si128(_mm_or_si128(dd_lo, dd_hi), xmm5); // r3<<1|r2>>62:r2<<1|r1>>63
162}
163
164static __m128i gcm_reduce(__m128i xx)
165{
166    //                                            // xx = x1:x0
167    /* [CLMUL-WP] Algorithm 5 Step 2 */
168    __m128i aa = _mm_slli_epi64(xx, 63);          // x1<<63:x0<<63 = stuff:a
169    __m128i bb = _mm_slli_epi64(xx, 62);          // x1<<62:x0<<62 = stuff:b
170    __m128i cc = _mm_slli_epi64(xx, 57);          // x1<<57:x0<<57 = stuff:c
171    __m128i dd = _mm_slli_si128(_mm_xor_si128(_mm_xor_si128(aa, bb), cc), 8); // a+b+c:0
172    return _mm_xor_si128(dd, xx);                 // x1+a+b+c:x0 = d:x0
173}
174
175static __m128i gcm_mix(__m128i dx)
176{
177    /* [CLMUL-WP] Algorithm 5 Steps 3 and 4 */
178    __m128i ee = _mm_srli_epi64(dx, 1);           // e1:x0>>1 = e1:e0'
179    __m128i ff = _mm_srli_epi64(dx, 2);           // f1:x0>>2 = f1:f0'
180    __m128i gg = _mm_srli_epi64(dx, 7);           // g1:x0>>7 = g1:g0'
181
182    // e0'+f0'+g0' is almost e0+f0+g0, except for some missing
183    // bits carried from d. Now get those bits back in.
184    __m128i eh = _mm_slli_epi64(dx, 63);          // d<<63:stuff
185    __m128i fh = _mm_slli_epi64(dx, 62);          // d<<62:stuff
186    __m128i gh = _mm_slli_epi64(dx, 57);          // d<<57:stuff
187    __m128i hh = _mm_srli_si128(_mm_xor_si128(_mm_xor_si128(eh, fh), gh), 8); // 0:missing bits of d
188
189    return _mm_xor_si128(_mm_xor_si128(_mm_xor_si128(_mm_xor_si128(ee, ff), gg), hh), dx);
190}
191
192void mbedtls_aesni_gcm_mult(unsigned char c[16],
193                            const unsigned char a[16],
194                            const unsigned char b[16])
195{
196    __m128i aa = { 0 }, bb = { 0 }, cc, dd;
197
198    /* The inputs are in big-endian order, so byte-reverse them */
199    for (size_t i = 0; i < 16; i++) {
200        ((uint8_t *) &aa)[i] = a[15 - i];
201        ((uint8_t *) &bb)[i] = b[15 - i];
202    }
203
204    gcm_clmul(aa, bb, &cc, &dd);
205    gcm_shift(&cc, &dd);
206    /*
207     * Now reduce modulo the GCM polynomial x^128 + x^7 + x^2 + x + 1
208     * using [CLMUL-WP] algorithm 5 (p. 18).
209     * Currently dd:cc holds x3:x2:x1:x0 (already shifted).
210     */
211    __m128i dx = gcm_reduce(cc);
212    __m128i xh = gcm_mix(dx);
213    cc = _mm_xor_si128(xh, dd); // x3+h1:x2+h0
214
215    /* Now byte-reverse the outputs */
216    for (size_t i = 0; i < 16; i++) {
217        c[i] = ((uint8_t *) &cc)[15 - i];
218    }
219
220    return;
221}
222
223/*
224 * Compute decryption round keys from encryption round keys
225 */
226#if !defined(MBEDTLS_BLOCK_CIPHER_NO_DECRYPT)
227void mbedtls_aesni_inverse_key(unsigned char *invkey,
228                               const unsigned char *fwdkey, int nr)
229{
230    __m128i *ik = (__m128i *) invkey;
231    const __m128i *fk = (const __m128i *) fwdkey + nr;
232
233    *ik = *fk;
234    for (--fk, ++ik; fk > (const __m128i *) fwdkey; --fk, ++ik) {
235        *ik = _mm_aesimc_si128(*fk);
236    }
237    *ik = *fk;
238}
239#endif
240
241/*
242 * Key expansion, 128-bit case
243 */
244static __m128i aesni_set_rk_128(__m128i state, __m128i xword)
245{
246    /*
247     * Finish generating the next round key.
248     *
249     * On entry state is r3:r2:r1:r0 and xword is X:stuff:stuff:stuff
250     * with X = rot( sub( r3 ) ) ^ RCON (obtained with AESKEYGENASSIST).
251     *
252     * On exit, xword is r7:r6:r5:r4
253     * with r4 = X + r0, r5 = r4 + r1, r6 = r5 + r2, r7 = r6 + r3
254     * and this is returned, to be written to the round key buffer.
255     */
256    xword = _mm_shuffle_epi32(xword, 0xff);   // X:X:X:X
257    xword = _mm_xor_si128(xword, state);      // X+r3:X+r2:X+r1:r4
258    state = _mm_slli_si128(state, 4);         // r2:r1:r0:0
259    xword = _mm_xor_si128(xword, state);      // X+r3+r2:X+r2+r1:r5:r4
260    state = _mm_slli_si128(state, 4);         // r1:r0:0:0
261    xword = _mm_xor_si128(xword, state);      // X+r3+r2+r1:r6:r5:r4
262    state = _mm_slli_si128(state, 4);         // r0:0:0:0
263    state = _mm_xor_si128(xword, state);      // r7:r6:r5:r4
264    return state;
265}
266
267static void aesni_setkey_enc_128(unsigned char *rk_bytes,
268                                 const unsigned char *key)
269{
270    __m128i *rk = (__m128i *) rk_bytes;
271
272    memcpy(&rk[0], key, 16);
273    rk[1] = aesni_set_rk_128(rk[0], _mm_aeskeygenassist_si128(rk[0], 0x01));
274    rk[2] = aesni_set_rk_128(rk[1], _mm_aeskeygenassist_si128(rk[1], 0x02));
275    rk[3] = aesni_set_rk_128(rk[2], _mm_aeskeygenassist_si128(rk[2], 0x04));
276    rk[4] = aesni_set_rk_128(rk[3], _mm_aeskeygenassist_si128(rk[3], 0x08));
277    rk[5] = aesni_set_rk_128(rk[4], _mm_aeskeygenassist_si128(rk[4], 0x10));
278    rk[6] = aesni_set_rk_128(rk[5], _mm_aeskeygenassist_si128(rk[5], 0x20));
279    rk[7] = aesni_set_rk_128(rk[6], _mm_aeskeygenassist_si128(rk[6], 0x40));
280    rk[8] = aesni_set_rk_128(rk[7], _mm_aeskeygenassist_si128(rk[7], 0x80));
281    rk[9] = aesni_set_rk_128(rk[8], _mm_aeskeygenassist_si128(rk[8], 0x1B));
282    rk[10] = aesni_set_rk_128(rk[9], _mm_aeskeygenassist_si128(rk[9], 0x36));
283}
284
285/*
286 * Key expansion, 192-bit case
287 */
288#if !defined(MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH)
289static void aesni_set_rk_192(__m128i *state0, __m128i *state1, __m128i xword,
290                             unsigned char *rk)
291{
292    /*
293     * Finish generating the next 6 quarter-keys.
294     *
295     * On entry state0 is r3:r2:r1:r0, state1 is stuff:stuff:r5:r4
296     * and xword is stuff:stuff:X:stuff with X = rot( sub( r3 ) ) ^ RCON
297     * (obtained with AESKEYGENASSIST).
298     *
299     * On exit, state0 is r9:r8:r7:r6 and state1 is stuff:stuff:r11:r10
300     * and those are written to the round key buffer.
301     */
302    xword = _mm_shuffle_epi32(xword, 0x55);   // X:X:X:X
303    xword = _mm_xor_si128(xword, *state0);    // X+r3:X+r2:X+r1:X+r0
304    *state0 = _mm_slli_si128(*state0, 4);     // r2:r1:r0:0
305    xword = _mm_xor_si128(xword, *state0);    // X+r3+r2:X+r2+r1:X+r1+r0:X+r0
306    *state0 = _mm_slli_si128(*state0, 4);     // r1:r0:0:0
307    xword = _mm_xor_si128(xword, *state0);    // X+r3+r2+r1:X+r2+r1+r0:X+r1+r0:X+r0
308    *state0 = _mm_slli_si128(*state0, 4);     // r0:0:0:0
309    xword = _mm_xor_si128(xword, *state0);    // X+r3+r2+r1+r0:X+r2+r1+r0:X+r1+r0:X+r0
310    *state0 = xword;                          // = r9:r8:r7:r6
311
312    xword = _mm_shuffle_epi32(xword, 0xff);   // r9:r9:r9:r9
313    xword = _mm_xor_si128(xword, *state1);    // stuff:stuff:r9+r5:r9+r4
314    *state1 = _mm_slli_si128(*state1, 4);     // stuff:stuff:r4:0
315    xword = _mm_xor_si128(xword, *state1);    // stuff:stuff:r9+r5+r4:r9+r4
316    *state1 = xword;                          // = stuff:stuff:r11:r10
317
318    /* Store state0 and the low half of state1 into rk, which is conceptually
319     * an array of 24-byte elements. Since 24 is not a multiple of 16,
320     * rk is not necessarily aligned so just `*rk = *state0` doesn't work. */
321    memcpy(rk, state0, 16);
322    memcpy(rk + 16, state1, 8);
323}
324
325static void aesni_setkey_enc_192(unsigned char *rk,
326                                 const unsigned char *key)
327{
328    /* First round: use original key */
329    memcpy(rk, key, 24);
330    /* aes.c guarantees that rk is aligned on a 16-byte boundary. */
331    __m128i state0 = ((__m128i *) rk)[0];
332    __m128i state1 = _mm_loadl_epi64(((__m128i *) rk) + 1);
333
334    aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x01), rk + 24 * 1);
335    aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x02), rk + 24 * 2);
336    aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x04), rk + 24 * 3);
337    aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x08), rk + 24 * 4);
338    aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x10), rk + 24 * 5);
339    aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x20), rk + 24 * 6);
340    aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x40), rk + 24 * 7);
341    aesni_set_rk_192(&state0, &state1, _mm_aeskeygenassist_si128(state1, 0x80), rk + 24 * 8);
342}
343#endif /* !MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH */
344
345/*
346 * Key expansion, 256-bit case
347 */
348#if !defined(MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH)
349static void aesni_set_rk_256(__m128i state0, __m128i state1, __m128i xword,
350                             __m128i *rk0, __m128i *rk1)
351{
352    /*
353     * Finish generating the next two round keys.
354     *
355     * On entry state0 is r3:r2:r1:r0, state1 is r7:r6:r5:r4 and
356     * xword is X:stuff:stuff:stuff with X = rot( sub( r7 )) ^ RCON
357     * (obtained with AESKEYGENASSIST).
358     *
359     * On exit, *rk0 is r11:r10:r9:r8 and *rk1 is r15:r14:r13:r12
360     */
361    xword = _mm_shuffle_epi32(xword, 0xff);
362    xword = _mm_xor_si128(xword, state0);
363    state0 = _mm_slli_si128(state0, 4);
364    xword = _mm_xor_si128(xword, state0);
365    state0 = _mm_slli_si128(state0, 4);
366    xword = _mm_xor_si128(xword, state0);
367    state0 = _mm_slli_si128(state0, 4);
368    state0 = _mm_xor_si128(state0, xword);
369    *rk0 = state0;
370
371    /* Set xword to stuff:Y:stuff:stuff with Y = subword( r11 )
372     * and proceed to generate next round key from there */
373    xword = _mm_aeskeygenassist_si128(state0, 0x00);
374    xword = _mm_shuffle_epi32(xword, 0xaa);
375    xword = _mm_xor_si128(xword, state1);
376    state1 = _mm_slli_si128(state1, 4);
377    xword = _mm_xor_si128(xword, state1);
378    state1 = _mm_slli_si128(state1, 4);
379    xword = _mm_xor_si128(xword, state1);
380    state1 = _mm_slli_si128(state1, 4);
381    state1 = _mm_xor_si128(state1, xword);
382    *rk1 = state1;
383}
384
385static void aesni_setkey_enc_256(unsigned char *rk_bytes,
386                                 const unsigned char *key)
387{
388    __m128i *rk = (__m128i *) rk_bytes;
389
390    memcpy(&rk[0], key, 16);
391    memcpy(&rk[1], key + 16, 16);
392
393    /*
394     * Main "loop" - Generating one more key than necessary,
395     * see definition of mbedtls_aes_context.buf
396     */
397    aesni_set_rk_256(rk[0], rk[1], _mm_aeskeygenassist_si128(rk[1], 0x01), &rk[2], &rk[3]);
398    aesni_set_rk_256(rk[2], rk[3], _mm_aeskeygenassist_si128(rk[3], 0x02), &rk[4], &rk[5]);
399    aesni_set_rk_256(rk[4], rk[5], _mm_aeskeygenassist_si128(rk[5], 0x04), &rk[6], &rk[7]);
400    aesni_set_rk_256(rk[6], rk[7], _mm_aeskeygenassist_si128(rk[7], 0x08), &rk[8], &rk[9]);
401    aesni_set_rk_256(rk[8], rk[9], _mm_aeskeygenassist_si128(rk[9], 0x10), &rk[10], &rk[11]);
402    aesni_set_rk_256(rk[10], rk[11], _mm_aeskeygenassist_si128(rk[11], 0x20), &rk[12], &rk[13]);
403    aesni_set_rk_256(rk[12], rk[13], _mm_aeskeygenassist_si128(rk[13], 0x40), &rk[14], &rk[15]);
404}
405#endif /* !MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH */
406
407#if defined(MBEDTLS_POP_TARGET_PRAGMA)
408#if defined(__clang__)
409#pragma clang attribute pop
410#elif defined(__GNUC__)
411#pragma GCC pop_options
412#endif
413#undef MBEDTLS_POP_TARGET_PRAGMA
414#endif
415
416#else /* MBEDTLS_AESNI_HAVE_CODE == 1 */
417
418#if defined(__has_feature)
419#if __has_feature(memory_sanitizer)
420#warning \
421    "MBEDTLS_AESNI_C is known to cause spurious error reports with some memory sanitizers as they do not understand the assembly code."
422#endif
423#endif
424
425/*
426 * Binutils needs to be at least 2.19 to support AES-NI instructions.
427 * Unfortunately, a lot of users have a lower version now (2014-04).
428 * Emit bytecode directly in order to support "old" version of gas.
429 *
430 * Opcodes from the Intel architecture reference manual, vol. 3.
431 * We always use registers, so we don't need prefixes for memory operands.
432 * Operand macros are in gas order (src, dst) as opposed to Intel order
433 * (dst, src) in order to blend better into the surrounding assembly code.
434 */
435#define AESDEC(regs)      ".byte 0x66,0x0F,0x38,0xDE," regs "\n\t"
436#define AESDECLAST(regs)  ".byte 0x66,0x0F,0x38,0xDF," regs "\n\t"
437#define AESENC(regs)      ".byte 0x66,0x0F,0x38,0xDC," regs "\n\t"
438#define AESENCLAST(regs)  ".byte 0x66,0x0F,0x38,0xDD," regs "\n\t"
439#define AESIMC(regs)      ".byte 0x66,0x0F,0x38,0xDB," regs "\n\t"
440#define AESKEYGENA(regs, imm)  ".byte 0x66,0x0F,0x3A,0xDF," regs "," imm "\n\t"
441#define PCLMULQDQ(regs, imm)   ".byte 0x66,0x0F,0x3A,0x44," regs "," imm "\n\t"
442
443#define xmm0_xmm0   "0xC0"
444#define xmm0_xmm1   "0xC8"
445#define xmm0_xmm2   "0xD0"
446#define xmm0_xmm3   "0xD8"
447#define xmm0_xmm4   "0xE0"
448#define xmm1_xmm0   "0xC1"
449#define xmm1_xmm2   "0xD1"
450
451/*
452 * AES-NI AES-ECB block en(de)cryption
453 */
454int mbedtls_aesni_crypt_ecb(mbedtls_aes_context *ctx,
455                            int mode,
456                            const unsigned char input[16],
457                            unsigned char output[16])
458{
459    asm ("movdqu    (%3), %%xmm0    \n\t" // load input
460         "movdqu    (%1), %%xmm1    \n\t" // load round key 0
461         "pxor      %%xmm1, %%xmm0  \n\t" // round 0
462         "add       $16, %1         \n\t" // point to next round key
463         "subl      $1, %0          \n\t" // normal rounds = nr - 1
464         "test      %2, %2          \n\t" // mode?
465         "jz        2f              \n\t" // 0 = decrypt
466
467         "1:                        \n\t" // encryption loop
468         "movdqu    (%1), %%xmm1    \n\t" // load round key
469         AESENC(xmm1_xmm0)                // do round
470         "add       $16, %1         \n\t" // point to next round key
471         "subl      $1, %0          \n\t" // loop
472         "jnz       1b              \n\t"
473         "movdqu    (%1), %%xmm1    \n\t" // load round key
474         AESENCLAST(xmm1_xmm0)            // last round
475#if !defined(MBEDTLS_BLOCK_CIPHER_NO_DECRYPT)
476         "jmp       3f              \n\t"
477
478         "2:                        \n\t" // decryption loop
479         "movdqu    (%1), %%xmm1    \n\t"
480         AESDEC(xmm1_xmm0)                // do round
481         "add       $16, %1         \n\t"
482         "subl      $1, %0          \n\t"
483         "jnz       2b              \n\t"
484         "movdqu    (%1), %%xmm1    \n\t" // load round key
485         AESDECLAST(xmm1_xmm0)            // last round
486#endif
487
488         "3:                        \n\t"
489         "movdqu    %%xmm0, (%4)    \n\t" // export output
490         :
491         : "r" (ctx->nr), "r" (ctx->buf + ctx->rk_offset), "r" (mode), "r" (input), "r" (output)
492         : "memory", "cc", "xmm0", "xmm1");
493
494
495    return 0;
496}
497
498/*
499 * GCM multiplication: c = a times b in GF(2^128)
500 * Based on [CLMUL-WP] algorithms 1 (with equation 27) and 5.
501 */
502void mbedtls_aesni_gcm_mult(unsigned char c[16],
503                            const unsigned char a[16],
504                            const unsigned char b[16])
505{
506    unsigned char aa[16], bb[16], cc[16];
507    size_t i;
508
509    /* The inputs are in big-endian order, so byte-reverse them */
510    for (i = 0; i < 16; i++) {
511        aa[i] = a[15 - i];
512        bb[i] = b[15 - i];
513    }
514
515    asm ("movdqu (%0), %%xmm0               \n\t" // a1:a0
516         "movdqu (%1), %%xmm1               \n\t" // b1:b0
517
518         /*
519          * Caryless multiplication xmm2:xmm1 = xmm0 * xmm1
520          * using [CLMUL-WP] algorithm 1 (p. 12).
521          */
522         "movdqa %%xmm1, %%xmm2             \n\t" // copy of b1:b0
523         "movdqa %%xmm1, %%xmm3             \n\t" // same
524         "movdqa %%xmm1, %%xmm4             \n\t" // same
525         PCLMULQDQ(xmm0_xmm1, "0x00")             // a0*b0 = c1:c0
526         PCLMULQDQ(xmm0_xmm2, "0x11")             // a1*b1 = d1:d0
527         PCLMULQDQ(xmm0_xmm3, "0x10")             // a0*b1 = e1:e0
528         PCLMULQDQ(xmm0_xmm4, "0x01")             // a1*b0 = f1:f0
529         "pxor %%xmm3, %%xmm4               \n\t" // e1+f1:e0+f0
530         "movdqa %%xmm4, %%xmm3             \n\t" // same
531         "psrldq $8, %%xmm4                 \n\t" // 0:e1+f1
532         "pslldq $8, %%xmm3                 \n\t" // e0+f0:0
533         "pxor %%xmm4, %%xmm2               \n\t" // d1:d0+e1+f1
534         "pxor %%xmm3, %%xmm1               \n\t" // c1+e0+f1:c0
535
536         /*
537          * Now shift the result one bit to the left,
538          * taking advantage of [CLMUL-WP] eq 27 (p. 18)
539          */
540         "movdqa %%xmm1, %%xmm3             \n\t" // r1:r0
541         "movdqa %%xmm2, %%xmm4             \n\t" // r3:r2
542         "psllq $1, %%xmm1                  \n\t" // r1<<1:r0<<1
543         "psllq $1, %%xmm2                  \n\t" // r3<<1:r2<<1
544         "psrlq $63, %%xmm3                 \n\t" // r1>>63:r0>>63
545         "psrlq $63, %%xmm4                 \n\t" // r3>>63:r2>>63
546         "movdqa %%xmm3, %%xmm5             \n\t" // r1>>63:r0>>63
547         "pslldq $8, %%xmm3                 \n\t" // r0>>63:0
548         "pslldq $8, %%xmm4                 \n\t" // r2>>63:0
549         "psrldq $8, %%xmm5                 \n\t" // 0:r1>>63
550         "por %%xmm3, %%xmm1                \n\t" // r1<<1|r0>>63:r0<<1
551         "por %%xmm4, %%xmm2                \n\t" // r3<<1|r2>>62:r2<<1
552         "por %%xmm5, %%xmm2                \n\t" // r3<<1|r2>>62:r2<<1|r1>>63
553
554         /*
555          * Now reduce modulo the GCM polynomial x^128 + x^7 + x^2 + x + 1
556          * using [CLMUL-WP] algorithm 5 (p. 18).
557          * Currently xmm2:xmm1 holds x3:x2:x1:x0 (already shifted).
558          */
559         /* Step 2 (1) */
560         "movdqa %%xmm1, %%xmm3             \n\t" // x1:x0
561         "movdqa %%xmm1, %%xmm4             \n\t" // same
562         "movdqa %%xmm1, %%xmm5             \n\t" // same
563         "psllq $63, %%xmm3                 \n\t" // x1<<63:x0<<63 = stuff:a
564         "psllq $62, %%xmm4                 \n\t" // x1<<62:x0<<62 = stuff:b
565         "psllq $57, %%xmm5                 \n\t" // x1<<57:x0<<57 = stuff:c
566
567         /* Step 2 (2) */
568         "pxor %%xmm4, %%xmm3               \n\t" // stuff:a+b
569         "pxor %%xmm5, %%xmm3               \n\t" // stuff:a+b+c
570         "pslldq $8, %%xmm3                 \n\t" // a+b+c:0
571         "pxor %%xmm3, %%xmm1               \n\t" // x1+a+b+c:x0 = d:x0
572
573         /* Steps 3 and 4 */
574         "movdqa %%xmm1,%%xmm0              \n\t" // d:x0
575         "movdqa %%xmm1,%%xmm4              \n\t" // same
576         "movdqa %%xmm1,%%xmm5              \n\t" // same
577         "psrlq $1, %%xmm0                  \n\t" // e1:x0>>1 = e1:e0'
578         "psrlq $2, %%xmm4                  \n\t" // f1:x0>>2 = f1:f0'
579         "psrlq $7, %%xmm5                  \n\t" // g1:x0>>7 = g1:g0'
580         "pxor %%xmm4, %%xmm0               \n\t" // e1+f1:e0'+f0'
581         "pxor %%xmm5, %%xmm0               \n\t" // e1+f1+g1:e0'+f0'+g0'
582         // e0'+f0'+g0' is almost e0+f0+g0, ex\tcept for some missing
583         // bits carried from d. Now get those\t bits back in.
584         "movdqa %%xmm1,%%xmm3              \n\t" // d:x0
585         "movdqa %%xmm1,%%xmm4              \n\t" // same
586         "movdqa %%xmm1,%%xmm5              \n\t" // same
587         "psllq $63, %%xmm3                 \n\t" // d<<63:stuff
588         "psllq $62, %%xmm4                 \n\t" // d<<62:stuff
589         "psllq $57, %%xmm5                 \n\t" // d<<57:stuff
590         "pxor %%xmm4, %%xmm3               \n\t" // d<<63+d<<62:stuff
591         "pxor %%xmm5, %%xmm3               \n\t" // missing bits of d:stuff
592         "psrldq $8, %%xmm3                 \n\t" // 0:missing bits of d
593         "pxor %%xmm3, %%xmm0               \n\t" // e1+f1+g1:e0+f0+g0
594         "pxor %%xmm1, %%xmm0               \n\t" // h1:h0
595         "pxor %%xmm2, %%xmm0               \n\t" // x3+h1:x2+h0
596
597         "movdqu %%xmm0, (%2)               \n\t" // done
598         :
599         : "r" (aa), "r" (bb), "r" (cc)
600         : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
601
602    /* Now byte-reverse the outputs */
603    for (i = 0; i < 16; i++) {
604        c[i] = cc[15 - i];
605    }
606
607    return;
608}
609
610/*
611 * Compute decryption round keys from encryption round keys
612 */
613#if !defined(MBEDTLS_BLOCK_CIPHER_NO_DECRYPT)
614void mbedtls_aesni_inverse_key(unsigned char *invkey,
615                               const unsigned char *fwdkey, int nr)
616{
617    unsigned char *ik = invkey;
618    const unsigned char *fk = fwdkey + 16 * nr;
619
620    memcpy(ik, fk, 16);
621
622    for (fk -= 16, ik += 16; fk > fwdkey; fk -= 16, ik += 16) {
623        asm ("movdqu (%0), %%xmm0       \n\t"
624             AESIMC(xmm0_xmm0)
625             "movdqu %%xmm0, (%1)       \n\t"
626             :
627             : "r" (fk), "r" (ik)
628             : "memory", "xmm0");
629    }
630
631    memcpy(ik, fk, 16);
632}
633#endif
634
635/*
636 * Key expansion, 128-bit case
637 */
638static void aesni_setkey_enc_128(unsigned char *rk,
639                                 const unsigned char *key)
640{
641    asm ("movdqu (%1), %%xmm0               \n\t" // copy the original key
642         "movdqu %%xmm0, (%0)               \n\t" // as round key 0
643         "jmp 2f                            \n\t" // skip auxiliary routine
644
645         /*
646          * Finish generating the next round key.
647          *
648          * On entry xmm0 is r3:r2:r1:r0 and xmm1 is X:stuff:stuff:stuff
649          * with X = rot( sub( r3 ) ) ^ RCON.
650          *
651          * On exit, xmm0 is r7:r6:r5:r4
652          * with r4 = X + r0, r5 = r4 + r1, r6 = r5 + r2, r7 = r6 + r3
653          * and those are written to the round key buffer.
654          */
655         "1:                                \n\t"
656         "pshufd $0xff, %%xmm1, %%xmm1      \n\t" // X:X:X:X
657         "pxor %%xmm0, %%xmm1               \n\t" // X+r3:X+r2:X+r1:r4
658         "pslldq $4, %%xmm0                 \n\t" // r2:r1:r0:0
659         "pxor %%xmm0, %%xmm1               \n\t" // X+r3+r2:X+r2+r1:r5:r4
660         "pslldq $4, %%xmm0                 \n\t" // etc
661         "pxor %%xmm0, %%xmm1               \n\t"
662         "pslldq $4, %%xmm0                 \n\t"
663         "pxor %%xmm1, %%xmm0               \n\t" // update xmm0 for next time!
664         "add $16, %0                       \n\t" // point to next round key
665         "movdqu %%xmm0, (%0)               \n\t" // write it
666         "ret                               \n\t"
667
668         /* Main "loop" */
669         "2:                                \n\t"
670         AESKEYGENA(xmm0_xmm1, "0x01")      "call 1b \n\t"
671         AESKEYGENA(xmm0_xmm1, "0x02")      "call 1b \n\t"
672         AESKEYGENA(xmm0_xmm1, "0x04")      "call 1b \n\t"
673         AESKEYGENA(xmm0_xmm1, "0x08")      "call 1b \n\t"
674         AESKEYGENA(xmm0_xmm1, "0x10")      "call 1b \n\t"
675         AESKEYGENA(xmm0_xmm1, "0x20")      "call 1b \n\t"
676         AESKEYGENA(xmm0_xmm1, "0x40")      "call 1b \n\t"
677         AESKEYGENA(xmm0_xmm1, "0x80")      "call 1b \n\t"
678         AESKEYGENA(xmm0_xmm1, "0x1B")      "call 1b \n\t"
679         AESKEYGENA(xmm0_xmm1, "0x36")      "call 1b \n\t"
680         :
681         : "r" (rk), "r" (key)
682         : "memory", "cc", "0");
683}
684
685/*
686 * Key expansion, 192-bit case
687 */
688#if !defined(MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH)
689static void aesni_setkey_enc_192(unsigned char *rk,
690                                 const unsigned char *key)
691{
692    asm ("movdqu (%1), %%xmm0   \n\t" // copy original round key
693         "movdqu %%xmm0, (%0)   \n\t"
694         "add $16, %0           \n\t"
695         "movq 16(%1), %%xmm1   \n\t"
696         "movq %%xmm1, (%0)     \n\t"
697         "add $8, %0            \n\t"
698         "jmp 2f                \n\t" // skip auxiliary routine
699
700         /*
701          * Finish generating the next 6 quarter-keys.
702          *
703          * On entry xmm0 is r3:r2:r1:r0, xmm1 is stuff:stuff:r5:r4
704          * and xmm2 is stuff:stuff:X:stuff with X = rot( sub( r3 ) ) ^ RCON.
705          *
706          * On exit, xmm0 is r9:r8:r7:r6 and xmm1 is stuff:stuff:r11:r10
707          * and those are written to the round key buffer.
708          */
709         "1:                            \n\t"
710         "pshufd $0x55, %%xmm2, %%xmm2  \n\t" // X:X:X:X
711         "pxor %%xmm0, %%xmm2           \n\t" // X+r3:X+r2:X+r1:r4
712         "pslldq $4, %%xmm0             \n\t" // etc
713         "pxor %%xmm0, %%xmm2           \n\t"
714         "pslldq $4, %%xmm0             \n\t"
715         "pxor %%xmm0, %%xmm2           \n\t"
716         "pslldq $4, %%xmm0             \n\t"
717         "pxor %%xmm2, %%xmm0           \n\t" // update xmm0 = r9:r8:r7:r6
718         "movdqu %%xmm0, (%0)           \n\t"
719         "add $16, %0                   \n\t"
720         "pshufd $0xff, %%xmm0, %%xmm2  \n\t" // r9:r9:r9:r9
721         "pxor %%xmm1, %%xmm2           \n\t" // stuff:stuff:r9+r5:r10
722         "pslldq $4, %%xmm1             \n\t" // r2:r1:r0:0
723         "pxor %%xmm2, %%xmm1           \n\t" // xmm1 = stuff:stuff:r11:r10
724         "movq %%xmm1, (%0)             \n\t"
725         "add $8, %0                    \n\t"
726         "ret                           \n\t"
727
728         "2:                            \n\t"
729         AESKEYGENA(xmm1_xmm2, "0x01")  "call 1b \n\t"
730         AESKEYGENA(xmm1_xmm2, "0x02")  "call 1b \n\t"
731         AESKEYGENA(xmm1_xmm2, "0x04")  "call 1b \n\t"
732         AESKEYGENA(xmm1_xmm2, "0x08")  "call 1b \n\t"
733         AESKEYGENA(xmm1_xmm2, "0x10")  "call 1b \n\t"
734         AESKEYGENA(xmm1_xmm2, "0x20")  "call 1b \n\t"
735         AESKEYGENA(xmm1_xmm2, "0x40")  "call 1b \n\t"
736         AESKEYGENA(xmm1_xmm2, "0x80")  "call 1b \n\t"
737
738         :
739         : "r" (rk), "r" (key)
740         : "memory", "cc", "0");
741}
742#endif /* !MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH */
743
744/*
745 * Key expansion, 256-bit case
746 */
747#if !defined(MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH)
748static void aesni_setkey_enc_256(unsigned char *rk,
749                                 const unsigned char *key)
750{
751    asm ("movdqu (%1), %%xmm0           \n\t"
752         "movdqu %%xmm0, (%0)           \n\t"
753         "add $16, %0                   \n\t"
754         "movdqu 16(%1), %%xmm1         \n\t"
755         "movdqu %%xmm1, (%0)           \n\t"
756         "jmp 2f                        \n\t" // skip auxiliary routine
757
758         /*
759          * Finish generating the next two round keys.
760          *
761          * On entry xmm0 is r3:r2:r1:r0, xmm1 is r7:r6:r5:r4 and
762          * xmm2 is X:stuff:stuff:stuff with X = rot( sub( r7 )) ^ RCON
763          *
764          * On exit, xmm0 is r11:r10:r9:r8 and xmm1 is r15:r14:r13:r12
765          * and those have been written to the output buffer.
766          */
767         "1:                                \n\t"
768         "pshufd $0xff, %%xmm2, %%xmm2      \n\t"
769         "pxor %%xmm0, %%xmm2               \n\t"
770         "pslldq $4, %%xmm0                 \n\t"
771         "pxor %%xmm0, %%xmm2               \n\t"
772         "pslldq $4, %%xmm0                 \n\t"
773         "pxor %%xmm0, %%xmm2               \n\t"
774         "pslldq $4, %%xmm0                 \n\t"
775         "pxor %%xmm2, %%xmm0               \n\t"
776         "add $16, %0                       \n\t"
777         "movdqu %%xmm0, (%0)               \n\t"
778
779         /* Set xmm2 to stuff:Y:stuff:stuff with Y = subword( r11 )
780          * and proceed to generate next round key from there */
781         AESKEYGENA(xmm0_xmm2, "0x00")
782         "pshufd $0xaa, %%xmm2, %%xmm2      \n\t"
783         "pxor %%xmm1, %%xmm2               \n\t"
784         "pslldq $4, %%xmm1                 \n\t"
785         "pxor %%xmm1, %%xmm2               \n\t"
786         "pslldq $4, %%xmm1                 \n\t"
787         "pxor %%xmm1, %%xmm2               \n\t"
788         "pslldq $4, %%xmm1                 \n\t"
789         "pxor %%xmm2, %%xmm1               \n\t"
790         "add $16, %0                       \n\t"
791         "movdqu %%xmm1, (%0)               \n\t"
792         "ret                               \n\t"
793
794         /*
795          * Main "loop" - Generating one more key than necessary,
796          * see definition of mbedtls_aes_context.buf
797          */
798         "2:                                \n\t"
799         AESKEYGENA(xmm1_xmm2, "0x01")      "call 1b \n\t"
800         AESKEYGENA(xmm1_xmm2, "0x02")      "call 1b \n\t"
801         AESKEYGENA(xmm1_xmm2, "0x04")      "call 1b \n\t"
802         AESKEYGENA(xmm1_xmm2, "0x08")      "call 1b \n\t"
803         AESKEYGENA(xmm1_xmm2, "0x10")      "call 1b \n\t"
804         AESKEYGENA(xmm1_xmm2, "0x20")      "call 1b \n\t"
805         AESKEYGENA(xmm1_xmm2, "0x40")      "call 1b \n\t"
806         :
807         : "r" (rk), "r" (key)
808         : "memory", "cc", "0");
809}
810#endif /* !MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH */
811
812#endif  /* MBEDTLS_AESNI_HAVE_CODE */
813
814/*
815 * Key expansion, wrapper
816 */
817int mbedtls_aesni_setkey_enc(unsigned char *rk,
818                             const unsigned char *key,
819                             size_t bits)
820{
821    switch (bits) {
822        case 128: aesni_setkey_enc_128(rk, key); break;
823#if !defined(MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH)
824        case 192: aesni_setkey_enc_192(rk, key); break;
825        case 256: aesni_setkey_enc_256(rk, key); break;
826#endif /* !MBEDTLS_AES_ONLY_128_BIT_KEY_LENGTH */
827        default: return MBEDTLS_ERR_AES_INVALID_KEY_LENGTH;
828    }
829
830    return 0;
831}
832
833#endif /* MBEDTLS_AESNI_HAVE_CODE */
834
835#endif /* MBEDTLS_AESNI_C */
836