rsa/rsa_pk1.c: remove memcpy calls from RSA_padding_check_PKCS1_type_2.

And make RSAErr call unconditional.
Reviewed-by: NRichard Levitte <levitte@openssl.org>
Reviewed-by: NMatt Caswell <matt@openssl.org>
(cherry picked from commit e875b0cf2f10bf2adf73e0c2ec81428290f4660c)

crypto/rsa/rsa_pk1.c

@@ -158,7 +158,7 @@ int RSA_padding_check_PKCS1_type_2(unsigned char *to, int tlen,
     int i;
     /* |em| is the encoded message, zero-padded to exactly |num| bytes */
     unsigned char *em = NULL;
-    unsigned int good, found_zero_byte;
+    unsigned int good, found_zero_byte, mask;
     int zero_index = 0, msg_index, mlen = -1;
 
     if (tlen < 0 || flen < 0)
@@ -169,39 +169,41 @@ int RSA_padding_check_PKCS1_type_2(unsigned char *to, int tlen,
      * section 7.2.2.
      */
 
-    if (flen > num)
-        goto err;
-
-    if (num < 11)
-        goto err;
+    if (flen > num || num < 11) {
+        RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2,
+               RSA_R_PKCS_DECODING_ERROR);
+        return -1;
+    }
 
-    if (flen != num) {
-        em = OPENSSL_zalloc(num);
-        if (em == NULL) {
-            RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2, ERR_R_MALLOC_FAILURE);
-            return -1;
-        }
-        /*
-         * Caller is encouraged to pass zero-padded message created with
-         * BN_bn2binpad, but if it doesn't, we do this zero-padding copy
-         * to avoid leaking that information. The copy still leaks some
-         * side-channel information, but it's impossible to have a fixed
-         * memory access pattern since we can't read out of the bounds of
-         * |from|.
-         */
-        memcpy(em + num - flen, from, flen);
-        from = em;
+    em = OPENSSL_malloc(num);
+    if (em == NULL) {
+        RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2, ERR_R_MALLOC_FAILURE);
+        return -1;
+    }
+    /*
+     * Caller is encouraged to pass zero-padded message created with
+     * BN_bn2binpad. Trouble is that since we can't read out of |from|'s
+     * bounds, it's impossible to have an invariant memory access pattern
+     * in case |from| was not zero-padded in advance.
+     */
+    for (from += flen, em += num, i = 0; i < num; i++) {
+        mask = ~constant_time_is_zero(flen);
+        flen -= 1 & mask;
+        from -= 1 & mask;
+        *--em = *from & mask;
     }
+    from = em;
 
     good = constant_time_is_zero(from[0]);
     good &= constant_time_eq(from[1], 2);
 
+    /* scan over padding data */
     found_zero_byte = 0;
     for (i = 2; i < num; i++) {
         unsigned int equals0 = constant_time_is_zero(from[i]);
-        zero_index =
-            constant_time_select_int(~found_zero_byte & equals0, i,
-                                     zero_index);
+
+        zero_index = constant_time_select_int(~found_zero_byte & equals0,
+                                              i, zero_index);
         found_zero_byte |= equals0;
     }
 
@@ -210,7 +212,7 @@ int RSA_padding_check_PKCS1_type_2(unsigned char *to, int tlen,
      * If we never found a 0-byte, then |zero_index| is 0 and the check
      * also fails.
      */
-    good &= constant_time_ge((unsigned int)(zero_index), 2 + 8);
+    good &= constant_time_ge(zero_index, 2 + 8);
 
     /*
      * Skip the zero byte. This is incorrect if we never found a zero-byte
@@ -220,27 +222,34 @@ int RSA_padding_check_PKCS1_type_2(unsigned char *to, int tlen,
     mlen = num - msg_index;
 
     /*
-     * For good measure, do this check in constant time as well; it could
-     * leak something if |tlen| was assuming valid padding.
+     * For good measure, do this check in constant time as well.
      */
-    good &= constant_time_ge((unsigned int)(tlen), (unsigned int)(mlen));
+    good &= constant_time_ge(tlen, mlen);
 
     /*
-     * We can't continue in constant-time because we need to copy the result
-     * and we cannot fake its length. This unavoidably leaks timing
-     * information at the API boundary.
+     * Even though we can't fake result's length, we can pretend copying
+     * |tlen| bytes where |mlen| bytes would be real. Last |tlen| of |num|
+     * bytes are viewed as circular buffer with start at |tlen|-|mlen'|,
+     * where |mlen'| is "saturated" |mlen| value. Deducing information
+     * about failure or |mlen| would take attacker's ability to observe
+     * memory access pattern with byte granularity *as it occurs*. It
+     * should be noted that failure is indistinguishable from normal
+     * operation if |tlen| is fixed by protocol.
      */
-    if (!good) {
-        mlen = -1;
-        goto err;
-    }
+    tlen = constant_time_select_int(constant_time_lt(num, tlen), num, tlen);
+    msg_index = constant_time_select_int(good, msg_index, num - tlen);
+    mlen = num - msg_index;
+    for (from += msg_index, mask = good, i = 0; i < tlen; i++) {
+        unsigned int equals = constant_time_eq(i, mlen);
 
-    memcpy(to, from + msg_index, mlen);
+        from -= tlen & equals;  /* if (i == mlen) rewind   */
+        mask &= mask ^ equals;  /* if (i == mlen) mask = 0 */
+        to[i] = constant_time_select_8(mask, from[i], to[i]);
+    }
 
- err:
     OPENSSL_clear_free(em, num);
-    if (mlen == -1)
-        RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2,
-               RSA_R_PKCS_DECODING_ERROR);
-    return mlen;
+    RSAerr(RSA_F_RSA_PADDING_CHECK_PKCS1_TYPE_2, RSA_R_PKCS_DECODING_ERROR);
+    err_clear_last_constant_time(1 & good);
+
+    return constant_time_select_int(good, mlen, -1);
 }

doc/man3/RSA_padding_add_PKCS1_type_1.pod

@@ -110,7 +110,12 @@ L<ERR_get_error(3)>.
 The RSA_padding_check_PKCS1_type_2() padding check leaks timing
 information which can potentially be used to mount a Bleichenbacher
 padding oracle attack. This is an inherent weakness in the PKCS #1
-v1.5 padding design. Prefer PKCS1_OAEP padding.
+v1.5 padding design. Prefer PKCS1_OAEP padding. Otherwise it can
+be recommended to pass zero-padded B<f>, so that B<fl> equals to
+B<rsa_len>, and if fixed by protocol, B<tlen> being set to the
+expected length. In such case leakage would be minimal, it would
+take attacker's ability to observe memory access pattern with byte
+granilarity as it occurs, post-factum timing analysis won't do.
 
 =head1 SEE ALSO