Undefined Pointer Arithmetic

In commits a004e72b9 (1.0.2) and 6f35f6deb (1.0.1) we released a fix for CVE-2016-2177. The fix corrects a common coding idiom present in OpenSSL 1.0.2 and OpenSSL 1.0.1 which actually relies on a usage of pointer arithmetic that is undefined in the C specification. The problem does not exist in master (OpenSSL 1.1.0) which refactored this code some while ago. This usage could give rise to a low severity security issue in certain unusual scenarios. The OpenSSL security policy (https://www.openssl.org/policies/secpolicy.html) states that we publish low severity issues directly to our public repository, and they get rolled up into the next release whenever that happens. The rest of this blog post describes the problem in a little more detail, explains the scenarios where a security issue could arise and why this issue has been rated as low severity.

The coding idiom we are talking about here is this one:

    if (p + len > limit)
    {
        return; /* Too long */
    }

Where p points to some malloc’d data of SIZE bytes and limit == p + SIZE. len here could be from some externally supplied data (e.g. from a TLS message).

The idea here is that we are performing a length check on peer supplied data to ensure that len (received from the peer) is actually within the bounds of the supplied data. The problem is that this pointer arithmetic is only defined by the C90 standard (which we aim to conform to) if len <= SIZE, so if len is too long then this is undefined behaviour, and of course (theoretically) anything could happen. In practice of course it usually works the way you expect it to work and there is no issue except in the case where an overflow occurs.

In order for an overflow to occur p + len would have to be sufficiently large to exceed the maximum addressable location. Recall that len here is received from the peer. However in all the instances that we fixed it represents either one byte or two bytes, i.e. its maximum value is 0xFFFF. Assuming a flat 32 bit address space this means that if p < 0xFFFF0000 then there is no issue, or putting it another way, approx. 0.0015% of all conceivable addresses are potentially “at risk”.

Most architectures do not have the heap at the edge of high memory. Typically, one finds:

    0: Reserved low-memory
    LOW:    TEXT
            DATA
            BSS
            HEAP
            BREAK
            ...
            mapped libraries and files
            STACK

And often (on systems with a unified kernel/user address space):

HIGH:
    kernel

I don’t claim to know much about memory lay outs so perhaps there are architectures out there that don’t conform to this. So, assuming that they exist, what would the impact be if p was sufficiently large that an overflow could occur?

There are two primary locations where a 2 byte len could be an issue (I will ignore 1 byte lens because it seems highly unlikely that we will ever run into a problem with those in reality). The first location is whilst reading in the ciphersuites in a ClientHello, and the second is whilst reading in the extensions in the ClientHello.

Here is the (pre-patch) ciphersuite code:

    n2s(p, i);

    if (i == 0) {
        al = SSL_AD_ILLEGAL_PARAMETER;
        SSLerr(SSL_F_SSL3_GET_CLIENT_HELLO, SSL_R_NO_CIPHERS_SPECIFIED);
        goto f_err;
    }

    /* i bytes of cipher data + 1 byte for compression length later */
    if ((p + i + 1) > (d + n)) {
        /* not enough data */
        al = SSL_AD_DECODE_ERROR;
        SSLerr(SSL_F_SSL3_GET_CLIENT_HELLO, SSL_R_LENGTH_MISMATCH);
        goto f_err;
    }
    if (ssl_bytes_to_cipher_list(s, p, i, &(ciphers)) == NULL) {
        goto err;
    }
    p += i;

Here i represents the two byte length read from the peer, p is the pointer into our buffer, and d + n is the end of the buffer. If `p + i

1 overflows then we will end up passing an excessively large ivalue tossl_bytes_to_cipher_list(). Analysing that function it can be seen that it will loop over all of the memory fromp` onwards interpreting it as ciphersuite data bytes and attempting to read its value. This is likely to cause a crash once we overflow (if not before).

The analysis of the extensions code is similar and more complicated, but the outcome is the same. It will loop over the out-of-bounds memory interpreting it as extension data and will eventually crash.

It seems that the only way the above two issues could be exploited is via a DoS.

The final consideration in all of this, is how much of the above is under the control of the attacker. Clearly len is, but not the value of p. In order to exploit this an attacker would have to first find a system which is likely to allocate at the very highest end of address space and then send a high number of requests through until one “got lucky” and a DoS results. A possible alternative approach is to send out many requests to many hosts attempting to find one that is vulnerable.

It appears that such attacks, whilst possible, are unlikely and would be difficult to achieve successfully in reality - with a worst case scenario being a DoS. Given the difficulty of the attack and the relatively low impact we rated this as a low severity issue.

With thanks to Guido Vranken who originally reported this issue.

OpenSSL Blog

Undefined Pointer Arithmetic

Comments