Implement AES in different modes of operation, using AES-NI and

patches/nss-intel-aes-windows.patch

+diff --git a/nss/lib/freebl/ctr.c b/nss/lib/freebl/ctr.c
+index 3a2f1a6..7807585 100644
+--- a/nss/lib/freebl/ctr.c
++++ b/nss/lib/freebl/ctr.c
+@@ -12,6 +12,11 @@
+ #include "pkcs11t.h"
+ #include "secerr.h"
+ 
++#ifdef USE_HW_AES
++#include "intel-aes.h"
++#include "rijndael.h"
++#endif
++
+ SECStatus
+ CTR_InitContext(CTRContext *ctr, void *context, freeblCipherFunc cipher,
+                const unsigned char *param, unsigned int blocksize)
+@@ -165,3 +170,56 @@ CTR_Update(CTRContext *ctr, unsigned char *outbuf,
+     *outlen += inlen;
+     return SECSuccess;
+ }
++
++#if defined(USE_HW_AES) && defined(_MSC_VER)
++SECStatus
++CTR_Update_HW_AES(CTRContext *ctr, unsigned char *outbuf,
++               unsigned int *outlen, unsigned int maxout,
++               const unsigned char *inbuf, unsigned int inlen,
++               unsigned int blocksize)
++{
++    unsigned int tmp;
++    SECStatus rv;
++
++    if (maxout < inlen) {
++       *outlen = inlen;
++       PORT_SetError(SEC_ERROR_OUTPUT_LEN);
++       return SECFailure;
++    }
++    *outlen = 0;
++    if (ctr->bufPtr != blocksize) {
++       unsigned int needed = PR_MIN(blocksize-ctr->bufPtr, inlen);
++       ctr_xor(outbuf, inbuf, ctr->buffer+ctr->bufPtr, needed);
++       ctr->bufPtr += needed;
++       outbuf += needed;
++       inbuf += needed;
++       *outlen += needed;
++       inlen -= needed;
++       if (inlen == 0) {
++           return SECSuccess;
++       }
++       PORT_Assert(ctr->bufPtr == blocksize);
++    }
++
++    intel_aes_ctr_worker(((AESContext*)(ctr->context))->Nr)(
++       ctr, outbuf, outlen, maxout, inbuf, inlen, blocksize);
++    *outlen += inlen & (-16);
++    outbuf += inlen & (-16);
++    inbuf += inlen & (-16);
++    inlen &= 16 - 1;
++
++    if (inlen == 0) {
++       return SECSuccess;
++    }
++    rv = (*ctr->cipher)(ctr->context, ctr->buffer, &tmp, blocksize,
++                       ctr->counter, blocksize, blocksize);
++    ctr_GetNextCtr(ctr->counter, ctr->counterBits, blocksize);
++    if (rv != SECSuccess) {
++       return SECFailure;
++    }
++    ctr_xor(outbuf, inbuf, ctr->buffer, inlen);
++    ctr->bufPtr = inlen;
++    *outlen += inlen;
++    return SECSuccess;
++}
++#endif
+diff --git a/nss/lib/freebl/ctr.h b/nss/lib/freebl/ctr.h
+index 69ef150..e7645a2 100644
+--- a/nss/lib/freebl/ctr.h
++++ b/nss/lib/freebl/ctr.h
+@@ -41,4 +41,11 @@ SECStatus CTR_Update(CTRContext *ctr, unsigned char *outbuf,
+                        const unsigned char *inbuf, unsigned int inlen,
+                        unsigned int blocksize);
+ 
++#ifdef USE_HW_AES
++SECStatus CTR_Update_HW_AES(CTRContext *ctr, unsigned char *outbuf,
++                       unsigned int *outlen, unsigned int maxout,
++                       const unsigned char *inbuf, unsigned int inlen,
++                       unsigned int blocksize);
++#endif
++
+ #endif
+diff --git a/nss/lib/freebl/intel-aes.h b/nss/lib/freebl/intel-aes.h
+index 1e18007..3b71e5f 100644
+--- a/nss/lib/freebl/intel-aes.h
++++ b/nss/lib/freebl/intel-aes.h
+@@ -33,6 +33,12 @@ SECStatus intel_aes_decrypt_cbc_128(AESContext *cx, unsigned char *output,
+                                    const unsigned char *input,
+                                    unsigned int inputLen,
+                                    unsigned int blocksize);
++SECStatus intel_aes_encrypt_ctr_128(CTRContext *cx, unsigned char *output,
++                                   unsigned int *outputLen,
++                                   unsigned int maxOutputLen,
++                                   const unsigned char *input,
++                                   unsigned int inputLen,
++                                   unsigned int blocksize);
+ SECStatus intel_aes_encrypt_ecb_192(AESContext *cx, unsigned char *output,
+                                    unsigned int *outputLen,
+                                    unsigned int maxOutputLen,
+@@ -57,6 +63,12 @@ SECStatus intel_aes_decrypt_cbc_192(AESContext *cx, unsigned char *output,
+                                    const unsigned char *input,
+                                    unsigned int inputLen,
+                                    unsigned int blocksize);
++SECStatus intel_aes_encrypt_ctr_192(CTRContext *cx, unsigned char *output,
++                                   unsigned int *outputLen,
++                                   unsigned int maxOutputLen,
++                                   const unsigned char *input,
++                                   unsigned int inputLen,
++                                   unsigned int blocksize);
+ SECStatus intel_aes_encrypt_ecb_256(AESContext *cx, unsigned char *output,
+                                    unsigned int *outputLen,
+                                    unsigned int maxOutputLen,
+@@ -81,6 +93,12 @@ SECStatus intel_aes_decrypt_cbc_256(AESContext *cx, unsigned char *output,
+                                    const unsigned char *input,
+                                    unsigned int inputLen,
+                                    unsigned int blocksize);
++SECStatus intel_aes_encrypt_ctr_256(CTRContext *cx, unsigned char *output,
++                                   unsigned int *outputLen,
++                                   unsigned int maxOutputLen,
++                                   const unsigned char *input,
++                                   unsigned int inputLen,
++                                   unsigned int blocksize);
+ 
+ 
+ #define intel_aes_ecb_worker(encrypt, keysize) \
+@@ -102,6 +120,11 @@ SECStatus intel_aes_decrypt_cbc_256(AESContext *cx, unsigned char *output,
+       (keysize) == 24 ? intel_aes_decrypt_cbc_192 :    \
+       intel_aes_decrypt_cbc_256))
+ 
++#define intel_aes_ctr_worker(nr) \
++   ((nr) == 10 ? intel_aes_encrypt_ctr_128 :   \
++    (nr) == 12 ? intel_aes_encrypt_ctr_192 :   \
++    intel_aes_encrypt_ctr_256)
++
+ 
+ #define intel_aes_init(encrypt, keysize) \
+   do {                                                         \
+diff --git a/nss/lib/freebl/intel-gcm-wrap.c b/nss/lib/freebl/intel-gcm-wrap.c
+index b2f6f5e..d49592f 100644
+--- a/nss/lib/freebl/intel-gcm-wrap.c
++++ b/nss/lib/freebl/intel-gcm-wrap.c
+@@ -3,7 +3,7 @@
+  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+ /* Copyright(c) 2013, Intel Corp. */
+ 
+-/* Wrapper funcions for Intel optimized implementation of AES-GCM */
++/* Wrapper functions for Intel optimized implementation of AES-GCM */
+ 
+ #ifdef USE_HW_AES
+ 
+@@ -24,12 +24,8 @@
+ #include "intel-gcm.h"
+ #include "rijndael.h"
+ 
+-#if defined(__INTEL_COMPILER)
+-#include <ia32intrin.h> 
+-#elif defined(__GNUC__)
+ #include <emmintrin.h>
+ #include <tmmintrin.h>
+-#endif
+ 
+ 
+ struct intel_AES_GCMContextStr{
+diff --git a/nss/lib/freebl/rijndael.c b/nss/lib/freebl/rijndael.c
+index 8bb8905..88f231e 100644
+--- a/nss/lib/freebl/rijndael.c
++++ b/nss/lib/freebl/rijndael.c
+@@ -1171,7 +1171,12 @@ AES_InitContext(AESContext *cx, const unsigned char *key, unsigned int keysize,
+        break;
+     case NSS_AES_CTR:
+        cx->worker_cx = CTR_CreateContext(cx, cx->worker, iv, blocksize);
+-       cx->worker = (freeblCipherFunc) CTR_Update ;
++#if defined(USE_HW_AES) && defined(_MSC_VER)
++       if (use_hw_aes)
++           cx->worker = (freeblCipherFunc) CTR_Update_HW_AES;
++       else
++#endif
++           cx->worker = (freeblCipherFunc) CTR_Update;
+        cx->destroy = (freeblDestroyFunc) CTR_DestroyContext;
+        cx->isBlock = PR_FALSE;
+        break;