Remove some dead code

[openssl.git] / crypto / evp / e_aes_cbc_hmac_sha256.c

diff --git a/crypto/evp/e_aes_cbc_hmac_sha256.c b/crypto/evp/e_aes_cbc_hmac_sha256.c
index b1c586e6fd96b8b46f30d2f17c8c19b9dc8f9d8d..aaa724a2d6efe03eb1adf0fd7e67a7fd0cc3534f 100644 (file)
--- a/crypto/evp/e_aes_cbc_hmac_sha256.c
+++ b/crypto/evp/e_aes_cbc_hmac_sha256.c
@@ -60,6 +60,7 @@
 # include <openssl/sha.h>
 # include <openssl/rand.h>
 # include "modes_lcl.h"
+# include "constant_time_locl.h"
 
 # ifndef EVP_CIPH_FLAG_AEAD_CIPHER
 #  define EVP_CIPH_FLAG_AEAD_CIPHER       0x200000
@@ -498,7 +499,18 @@ static int aesni_cbc_hmac_sha256_cipher(EVP_CIPHER_CTX *ctx,
             iv = AES_BLOCK_SIZE;
 
 #  if defined(STITCHED_CALL)
+        /*
+         * Assembly stitch handles AVX-capable processors, but its
+         * performance is not optimal on AMD Jaguar, ~40% worse, for
+         * unknown reasons. Incidentally processor in question supports
+         * AVX, but not AMD-specific XOP extension, which can be used
+         * to identify it and avoid stitch invocation. So that after we
+         * establish that current CPU supports AVX, we even see if it's
+         * either even XOP-capable Bulldozer-based or GenuineIntel one.
+         */
         if (OPENSSL_ia32cap_P[1] & (1 << (60 - 32)) && /* AVX? */
+            ((OPENSSL_ia32cap_P[1] & (1 << (43 - 32))) /* XOP? */
+             | (OPENSSL_ia32cap_P[0] & (1<<30))) &&    /* "Intel CPU"? */
             plen > (sha_off + iv) &&
             (blocks = (plen - (sha_off + iv)) / SHA256_CBLOCK)) {
             SHA256_Update(&key->md, in + iv, sha_off);
@@ -578,10 +590,17 @@ static int aesni_cbc_hmac_sha256_cipher(EVP_CIPHER_CTX *ctx,
             maxpad |= (255 - maxpad) >> (sizeof(maxpad) * 8 - 8);
             maxpad &= 255;
 
+            mask = constant_time_ge(maxpad, pad);
+            ret &= mask;
+            /*
+             * If pad is invalid then we will fail the above test but we must
+             * continue anyway because we are in constant time code. However,
+             * we'll use the maxpad value instead of the supplied pad to make
+             * sure we perform well defined pointer arithmetic.
+             */
+            pad = constant_time_select(mask, pad, maxpad);
+
             inp_len = len - (SHA256_DIGEST_LENGTH + pad + 1);
-            mask = (0 - ((inp_len - len) >> (sizeof(inp_len) * 8 - 1)));
-            inp_len &= mask;
-            ret &= (int)mask;
 
             key->aux.tls_aad[plen - 2] = inp_len >> 8;
             key->aux.tls_aad[plen - 1] = inp_len;
@@ -811,7 +830,7 @@ static int aesni_cbc_hmac_sha256_ctrl(EVP_CIPHER_CTX *ctx, int type, int arg,
     case EVP_CTRL_AEAD_TLS1_AAD:
         {
             unsigned char *p = ptr;
-            unsigned int len = p[arg - 2] << 8 | p[arg - 1];
+            unsigned int len;
 
             if (arg != EVP_AEAD_TLS1_AAD_LEN)
                 return -1;
@@ -822,6 +841,8 @@ static int aesni_cbc_hmac_sha256_ctrl(EVP_CIPHER_CTX *ctx, int type, int arg,
                 key->payload_length = len;
                 if ((key->aux.tls_ver =
                      p[arg - 4] << 8 | p[arg - 3]) >= TLS1_1_VERSION) {
+                    if (len < AES_BLOCK_SIZE)
+                        return 0;
                     len -= AES_BLOCK_SIZE;
                     p[arg - 2] = len >> 8;
                     p[arg - 1] = len;