modes/ctr128.c: pay attention to ecount_buf alignment in CRYPTO_ctr128_encrypt.

It's never problem if CRYPTO_ctr128_encrypt is called from EVP, because
buffer in question is always aligned within EVP_CIPHER_CTX structure.

RT#4218

Reviewed-by: Richard Levitte <levitte@openssl.org>
(cherry picked from commit 5e4bbeb49f)

crypto/modes/ctr128.c

@@ -67,23 +67,20 @@
 /* increment counter (128-bit int) by 1 */
 static void ctr128_inc(unsigned char *counter)
 {
-    u32 n = 16;
-    u8 c;
+    u32 n = 16, c = 1;
 
     do {
         --n;
-        c = counter[n];
-        ++c;
-        counter[n] = c;
-        if (c)
-            return;
+        c += counter[n];
+        counter[n] = (u8)c;
+        c >>= 8;
     } while (n);
 }
 
 #if !defined(OPENSSL_SMALL_FOOTPRINT)
 static void ctr128_inc_aligned(unsigned char *counter)
 {
-    size_t *data, c, n;
+    size_t *data, c, d, n;
     const union {
         long one;
         char little;
@@ -91,20 +88,19 @@ static void ctr128_inc_aligned(unsigned char *counter)
         1
     };
 
-    if (is_endian.little) {
+    if (is_endian.little || ((size_t)counter % sizeof(size_t)) != 0) {
         ctr128_inc(counter);
         return;
     }
 
     data = (size_t *)counter;
+    c = 1;
     n = 16 / sizeof(size_t);
     do {
         --n;
-        c = data[n];
-        ++c;
-        data[n] = c;
-        if (c)
-            return;
+        d = data[n] += c;
+        /* did addition carry? */
+        c = ((d - c) ^ d) >> (sizeof(size_t) * 8 - 1);
     } while (n);
 }
 #endif
@@ -144,14 +140,14 @@ void CRYPTO_ctr128_encrypt(const unsigned char *in, unsigned char *out,
             }
 
 # if defined(STRICT_ALIGNMENT)
-            if (((size_t)in | (size_t)out | (size_t)ivec) % sizeof(size_t) !=
-                0)
+            if (((size_t)in | (size_t)out | (size_t)ecount_buf)
+                % sizeof(size_t) != 0)
                 break;
 # endif
             while (len >= 16) {
                 (*block) (ivec, ecount_buf, key);
                 ctr128_inc_aligned(ivec);
-                for (; n < 16; n += sizeof(size_t))
+                for (n = 0; n < 16; n += sizeof(size_t))
                     *(size_t *)(out + n) =
                         *(size_t *)(in + n) ^ *(size_t *)(ecount_buf + n);
                 len -= 16;
@@ -189,16 +185,13 @@ void CRYPTO_ctr128_encrypt(const unsigned char *in, unsigned char *out,
 /* increment upper 96 bits of 128-bit counter by 1 */
 static void ctr96_inc(unsigned char *counter)
 {
-    u32 n = 12;
-    u8 c;
+    u32 n = 12, c = 1;
 
     do {
         --n;
-        c = counter[n];
-        ++c;
-        counter[n] = c;
-        if (c)
-            return;
+        c += counter[n];
+        counter[n] = (u8)c;
+        c >>= 8;
     } while (n);
 }