From eeb11a6e510ef3fb687d0fc314546a5fc779bd98 Mon Sep 17 00:00:00 2001
From: toddouska <todd@yassl.com>
Date: Mon, 7 Jan 2013 14:06:58 -0800
Subject: [PATCH] fix rabbit and hc128 CTaoCrypt test buffers for aligned
 access only, allow TLS on intel w/o aligned stream buffers, otherwise align

---
 ctaocrypt/test/test.c       | 47 ++++++++++++++++++++++---------------
 cyassl/ctaocrypt/settings.h |  7 ++++++
 cyassl/ctaocrypt/types.h    |  3 ++-
 src/internal.c              | 24 +++++++++++++++++++
 4 files changed, 61 insertions(+), 20 deletions(-)

diff --git a/ctaocrypt/test/test.c b/ctaocrypt/test/test.c
index 08842fd30..1f4abd4da 100644
--- a/ctaocrypt/test/test.c
+++ b/ctaocrypt/test/test.c
@@ -1194,11 +1194,16 @@ int hc128_test(void)
         HC128 enc;
         HC128 dec;
 
-        Hc128_SetKey(&enc, (byte*)keys[i], (byte*)ivs[i]);
-        Hc128_SetKey(&dec, (byte*)keys[i], (byte*)ivs[i]);
+        /* align keys/ivs in plain/cipher buffers */
+        memcpy(plain,  keys[i], 16); 
+        memcpy(cipher, ivs[i],  16); 
 
-        Hc128_Process(&enc, cipher, (byte*)test_hc128[i].input,
-                    (word32)test_hc128[i].outLen);
+        Hc128_SetKey(&enc, plain, cipher);
+        Hc128_SetKey(&dec, plain, cipher);
+
+        /* align input */
+        memcpy(plain, test_hc128[i].input, test_hc128[i].outLen);
+        Hc128_Process(&enc, cipher, plain,  (word32)test_hc128[i].outLen);
         Hc128_Process(&dec, plain,  cipher, (word32)test_hc128[i].outLen);
 
         if (memcmp(plain, test_hc128[i].input, test_hc128[i].outLen))
@@ -1219,22 +1224,17 @@ int rabbit_test(void)
     byte cipher[16];
     byte plain[16];
 
-    const char* keys[] =   /* align with 3 extra bytes cause null is added */
+    const char* keys[] = 
     {           
-        "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00"
-            "\x00\x00\x00",
-        "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00"
-            "\x00\x00\x00",
+        "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00",
+        "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00",
         "\xAC\xC3\x51\xDC\xF1\x62\xFC\x3B\xFE\x36\x3D\x2E\x29\x13\x28\x91"
-            "\x00\x00\x00"
     };
 
-    const char* ivs[] = /* align with 3 extra bytes casue null is added */
+    const char* ivs[] =
     {
-        "\x00\x00\x00\x00\x00\x00\x00\x00"
-            "\x00\x00\x00",
-        "\x59\x7E\x26\xC1\x75\xF5\x73\xC3"
-            "\x00\x00\x00",
+        "\x00\x00\x00\x00\x00\x00\x00\x00",
+        "\x59\x7E\x26\xC1\x75\xF5\x73\xC3",
         0
     };
 
@@ -1265,12 +1265,21 @@ int rabbit_test(void)
     for (i = 0; i < times; ++i) {
         Rabbit enc;
         Rabbit dec;
+        byte*  iv;
 
-        RabbitSetKey(&enc, (byte*)keys[i], (byte*)ivs[i]);
-        RabbitSetKey(&dec, (byte*)keys[i], (byte*)ivs[i]);
+        /* align keys/ivs in plain/cipher buffers */
+        memcpy(plain,  keys[i], 16);
+        if (ivs[i]) {
+            memcpy(cipher, ivs[i],   8);
+            iv = cipher;
+        } else
+            iv = NULL;
+        RabbitSetKey(&enc, plain, iv);
+        RabbitSetKey(&dec, plain, iv);
 
-        RabbitProcess(&enc, cipher, (byte*)test_rabbit[i].input,
-                    (word32)test_rabbit[i].outLen);
+        /* align input */
+        memcpy(plain, test_rabbit[i].input, test_rabbit[i].outLen);
+        RabbitProcess(&enc, cipher, plain,  (word32)test_rabbit[i].outLen);
         RabbitProcess(&dec, plain,  cipher, (word32)test_rabbit[i].outLen);
 
         if (memcmp(plain, test_rabbit[i].input, test_rabbit[i].outLen))
diff --git a/cyassl/ctaocrypt/settings.h b/cyassl/ctaocrypt/settings.h
index c8075760e..2890ebd18 100644
--- a/cyassl/ctaocrypt/settings.h
+++ b/cyassl/ctaocrypt/settings.h
@@ -72,6 +72,13 @@
 
 #include <cyassl/ctaocrypt/visibility.h>
 
+/* stream ciphers except arc4 need 32bit alignment, intel ok without */
+#if defined(__x86_64__) || defined(__ia64__) || defined(__i386__)
+    #define NO_XSTREAM_ALIGNMENT
+#else
+    #define XSTREAM_ALIGNMENT
+#endif
+
 #ifdef IPHONE
     #define SIZEOF_LONG_LONG 8
 #endif
diff --git a/cyassl/ctaocrypt/types.h b/cyassl/ctaocrypt/types.h
index 89a5a2857..82bbaec5b 100644
--- a/cyassl/ctaocrypt/types.h
+++ b/cyassl/ctaocrypt/types.h
@@ -227,7 +227,8 @@ enum {
     DYNAMIC_TYPE_DTLS_POOL    = 34,
     DYNAMIC_TYPE_SOCKADDR     = 35,
     DYNAMIC_TYPE_LIBZ         = 36,
-    DYNAMIC_TYPE_ECC          = 37
+    DYNAMIC_TYPE_ECC          = 37,
+    DYNAMIC_TYPE_TMP_BUFFER   = 38
 };
 
 /* stack protection */
diff --git a/src/internal.c b/src/internal.c
index 4a79e6162..eef89380c 100644
--- a/src/internal.c
+++ b/src/internal.c
@@ -3000,12 +3000,36 @@ static INLINE int Encrypt(CYASSL* ssl, byte* out, const byte* input, word32 sz)
 
         #ifdef HAVE_HC128
             case hc128:
+                #ifdef XSTREAM_ALIGNMENT
+                if ((word)input % 4) {
+                    byte* tmp = (byte*)XMALLOC(sz, ssl->heap,
+                                               DYNAMIC_TYPE_TMP_BUFFER);
+                    if (tmp == NULL) return MEMORY_E;
+                    XMEMCPY(tmp, input, sz);
+                    Hc128_Process(ssl->encrypt.hc128, tmp, tmp, sz);
+                    XMEMCPY(out, tmp, sz);
+                    XFREE(tmp, ssl->heap, DYNAMIC_TYPE_TMP_BUFFER);
+                    break;
+                }
+                #endif
                 Hc128_Process(ssl->encrypt.hc128, out, input, sz);
                 break;
         #endif
 
         #ifdef BUILD_RABBIT
             case rabbit:
+                #ifdef XSTREAM_ALIGNMENT
+                if ((word)input % 4) {
+                    byte* tmp = (byte*)XMALLOC(sz, ssl->heap,
+                                               DYNAMIC_TYPE_TMP_BUFFER);
+                    if (tmp == NULL) return MEMORY_E;
+                    XMEMCPY(tmp, input, sz);
+                    RabbitProcess(ssl->encrypt.rabbit, tmp, tmp, sz);
+                    XMEMCPY(out, tmp, sz);
+                    XFREE(tmp, ssl->heap, DYNAMIC_TYPE_TMP_BUFFER);
+                    break;
+                }
+                #endif
                 RabbitProcess(ssl->encrypt.rabbit, out, input, sz);
                 break;
         #endif