Follow on from CVE-2016-2179

The investigation and analysis of CVE-2016-2179 highlighted a related flaw.

This commit fixes a security "near miss" in the buffered message handling
code. Ultimately this is not currently believed to be exploitable due to
the reasons outlined below, and therefore there is no CVE for this on its
own.

The issue this commit fixes is a MITM attack where the attacker can inject
a Finished message into the handshake. In the description below it is
assumed that the attacker injects the Finished message for the server to
receive it. The attack could work equally well the other way around (i.e
where the client receives the injected Finished message).

The MITM requires the following capabilities:
- The ability to manipulate the MTU that the client selects such that it
is small enough for the client to fragment Finished messages.
- The ability to selectively drop and modify records sent from the client
- The ability to inject its own records and send them to the server

The MITM forces the client to select a small MTU such that the client
will fragment the Finished message. Ideally for the attacker the first
fragment will contain all but the last byte of the Finished message,
with the second fragment containing the final byte.

During the handshake and prior to the client sending the CCS the MITM
injects a plaintext Finished message fragment to the server containing
all but the final byte of the Finished message. The message sequence
number should be the one expected to be used for the real Finished message.

OpenSSL will recognise that the received fragment is for the future and
will buffer it for later use.

After the client sends the CCS it then sends its own Finished message in
two fragments. The MITM causes the first of these fragments to be
dropped. The OpenSSL server will then receive the second of the fragments
and reassemble the complete Finished message consisting of the MITM
fragment and the final byte from the real client.

The advantage to the attacker in injecting a Finished message is that
this provides the capability to modify other handshake messages (e.g.
the ClientHello) undetected. A difficulty for the attacker is knowing in
advance what impact any of those changes might have on the final byte of
the handshake hash that is going to be sent in the "real" Finished
message. In the worst case for the attacker this means that only 1 in
256 of such injection attempts will succeed.

It may be possible in some situations for the attacker to improve this such
that all attempts succeed. For example if the handshake includes client
authentication then the final message flight sent by the client will
include a Certificate. Certificates are ASN.1 objects where the signed
portion is DER encoded. The non-signed portion could be BER encoded and so
the attacker could re-encode the certificate such that the hash for the
whole handshake comes to a different value. The certificate re-encoding
would not be detectable because only the non-signed portion is changed. As
this is the final flight of messages sent from the client the attacker
knows what the complete hanshake hash value will be that the client will
send - and therefore knows what the final byte will be. Through a process
of trial and error the attacker can re-encode the certificate until the
modified handhshake also has a hash with the same final byte. This means
that when the Finished message is verified by the server it will be
correct in all cases.

In practice the MITM would need to be able to perform the same attack
against both the client and the server. If the attack is only performed
against the server (say) then the server will not detect the modified
handshake, but the client will and will abort the connection.
Fortunately, although OpenSSL is vulnerable to Finished message
injection, it is not vulnerable if *both* client and server are OpenSSL.
The reason is that OpenSSL has a hard "floor" for a minimum MTU size
that it will never go below. This minimum means that a Finished message
will never be sent in a fragmented form and therefore the MITM does not
have one of its pre-requisites. Therefore this could only be exploited
if using OpenSSL and some other DTLS peer that had its own and separate
Finished message injection flaw.

The fix is to ensure buffered messages are cleared on epoch change.

Reviewed-by: Richard Levitte <levitte@openssl.org>

DTLS can handle out of order record delivery. Additionally since
handshake messages can be bigger than will fit into a single packet, the
messages can be fragmented across multiple records (as with normal TLS).
That means that the messages can arrive mixed up, and we have to
reassemble them. We keep a queue of buffered messages that are "from the
future", i.e. messages we're not ready to deal with yet but have arrived
early. The messages held there may not be full yet - they could be one
or more fragments that are still in the process of being reassembled.

The code assumes that we will eventually complete the reassembly and
when that occurs the complete message is removed from the queue at the
point that we need to use it.

However, DTLS is also tolerant of packet loss. To get around that DTLS
messages can be retransmitted. If we receive a full (non-fragmented)
message from the peer after previously having received a fragment of
that message, then we ignore the message in the queue and just use the
non-fragmented version. At that point the queued message will never get
removed.

Additionally the peer could send "future" messages that we never get to
in order to complete the handshake. Each message has a sequence number
(starting from 0). We will accept a message fragment for the current
message sequence number, or for any sequence up to 10 into the future.
However if the Finished message has a sequence number of 2, anything
greater than that in the queue is just left there.

So, in those two ways we can end up with "orphaned" data in the queue
that will never get removed - except when the connection is closed. At
that point all the queues are flushed.

An attacker could seek to exploit this by filling up the queues with
lots of large messages that are never going to be used in order to
attempt a DoS by memory exhaustion.

I will assume that we are only concerned with servers here. It does not
seem reasonable to be concerned about a memory exhaustion attack on a
client. They are unlikely to process enough connections for this to be
an issue.

A "long" handshake with many messages might be 5 messages long (in the
incoming direction), e.g. ClientHello, Certificate, ClientKeyExchange,
CertificateVerify, Finished. So this would be message sequence numbers 0
to 4. Additionally we can buffer up to 10 messages in the future.
Therefore the maximum number of messages that an attacker could send
that could get orphaned would typically be 15.

The maximum size that a DTLS message is allowed to be is defined by
max_cert_list, which by default is 100k. Therefore the maximum amount of
"orphaned" memory per connection is 1500k.

Message sequence numbers get reset after the Finished message, so
renegotiation will not extend the maximum number of messages that can be
orphaned per connection.

As noted above, the queues do get cleared when the connection is closed.
Therefore in order to mount an effective attack, an attacker would have
to open many simultaneous connections.

Issue reported by Quan Luo.

CVE-2016-2179

Reviewed-by: Richard Levitte <levitte@openssl.org>

The enable-zlib option was broken by the recent "const" changes.

Reviewed-by: Stephen Henson <steve@openssl.org>

Most of the time, this isn't strictly needed.  However, in the default
extern model (called relaxed refdef), symbols are treated as weak
common objects unless they are initialised.  The librarian doesn't
include weak symbols in the (static) libraries, which renders them
invisible when linking a program with said those libraries, which is a
problem at times.

Using the strict refdef model is much more like standard C on all
other platforms, and thereby avoid the issues that come with the
relaxed refdef model.

Reviewed-by: Rich Salz <rsalz@openssl.org>

RT#4628

Reviewed-by: Rich Salz <rsalz@openssl.org>

RT#4628

Reviewed-by: Rich Salz <rsalz@openssl.org>

Thanks to Brian Smith for reporting this.

Reviewed-by: Rich Salz <rsalz@openssl.org>

Reviewed-by: Richard Levitte <levitte@openssl.org>

Reviewed-by: Viktor Dukhovni <viktor@openssl.org>

Add mutable versions of X509_get0_notBefore and X509_get0_notAfter.

Rename X509_SIG_get0_mutable to X509_SIG_getm.

Reviewed-by: Viktor Dukhovni <viktor@openssl.org>

Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1475)

Reviewed-by: Rich Salz <rsalz@openssl.org>

MR: #3176

Reviewed-by: Dr. Stephen Henson <steve@openssl.org>

Reviewed-by: Rich Salz <rsalz@openssl.org>

Reviewed-by: Rich Salz <rsalz@openssl.org>

Reviewed-by: Rich Salz <rsalz@openssl.org>

Update certificate and CRL time routines to match new standard.

Reviewed-by: Rich Salz <rsalz@openssl.org>

The DANE API supports a DANE_FLAG_NO_DANE_EE_NAMECHECKS option, but
there was no way to exercise/enable it via s_client.  This commit
addresses that gap.

Reviewed-by: Rich Salz <rsalz@openssl.org>

Reviewed-by: Rich Salz <rsalz@openssl.org>

The certificate and CRL time setting functions used similar code,
combine into a single utility function.

Reviewed-by: Rich Salz <rsalz@openssl.org>

Reviewed-by: Richard Levitte <levitte@openssl.org>

Reviewed-by: Rich Salz <rsalz@openssl.org>

... without any interruption.

Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1468)

Reviewed-by: Richard Levitte <levitte@openssl.org>

Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Stephen Henson <steve@openssl.org>

A function error code needed updating due to merge issues.

Reviewed-by: Richard Levitte <levitte@openssl.org>

Clang was complaining about some unused functions. Moving the stack
declaration to the header seems to sort it. Also the certstatus variable
in dtlstest needed to be declared static.

Reviewed-by: Richard Levitte <levitte@openssl.org>

The DTLS implementation provides some protection against replay attacks
in accordance with RFC6347 section 4.1.2.6.

A sliding "window" of valid record sequence numbers is maintained with
the "right" hand edge of the window set to the highest sequence number we
have received so far. Records that arrive that are off the "left" hand
edge of the window are rejected. Records within the window are checked
against a list of records received so far. If we already received it then
we also reject the new record.

If we have not already received the record, or the sequence number is off
the right hand edge of the window then we verify the MAC of the record.
If MAC verification fails then we discard the record. Otherwise we mark
the record as received. If the sequence number was off the right hand edge
of the window, then we slide the window along so that the right hand edge
is in line with the newly received sequence number.

Records may arrive for future epochs, i.e. a record from after a CCS being
sent, can arrive before the CCS does if the packets get re-ordered. As we
have not yet received the CCS we are not yet in a position to decrypt or
validate the MAC of those records. OpenSSL places those records on an
unprocessed records queue. It additionally updates the window immediately,
even though we have not yet verified the MAC. This will only occur if
currently in a handshake/renegotiation.

This could be exploited by an attacker by sending a record for the next
epoch (which does not have to decrypt or have a valid MAC), with a very
large sequence number. This means the right hand edge of the window is
moved very far to the right, and all subsequent legitimate packets are
dropped causing a denial of service.

A similar effect can be achieved during the initial handshake. In this
case there is no MAC key negotiated yet. Therefore an attacker can send a
message for the current epoch with a very large sequence number. The code
will process the record as normal. If the hanshake message sequence number
(as opposed to the record sequence number that we have been talking about
so far) is in the future then the injected message is bufferred to be
handled later, but the window is still updated. Therefore all subsequent
legitimate handshake records are dropped. This aspect is not considered a
security issue because there are many ways for an attacker to disrupt the
initial handshake and prevent it from completing successfully (e.g.
injection of a handshake message will cause the Finished MAC to fail and
the handshake to be aborted). This issue comes about as a result of trying
to do replay protection, but having no integrity mechanism in place yet.
Does it even make sense to have replay protection in epoch 0? That
issue isn't addressed here though.

This addressed an OCAP Audit issue.

CVE-2016-2181

Reviewed-by: Richard Levitte <levitte@openssl.org>

Injects a record from epoch 1 during epoch 0 handshake, with a record
sequence number in the future, to test that the record replay protection
feature works as expected. This is described more fully in the next commit.

Reviewed-by: Richard Levitte <levitte@openssl.org>

During a DTLS handshake we may get records destined for the next epoch
arrive before we have processed the CCS. In that case we can't decrypt or
verify the record yet, so we buffer it for later use. When we do receive
the CCS we work through the queue of unprocessed records and process them.

Unfortunately the act of processing wipes out any existing packet data
that we were still working through. This includes any records from the new
epoch that were in the same packet as the CCS. We should only process the
buffered records if we've not got any data left.

Reviewed-by: Richard Levitte <levitte@openssl.org>

Add a test to inject a record from the next epoch during the handshake and
make sure it doesn't get processed immediately.

Reviewed-by: Richard Levitte <levitte@openssl.org>

Split the create_ssl_connection() helper function into two steps: one to
create the SSL objects, and one to actually create the connection. This
provides the ability to make changes to the SSL object before the
connection is actually made.

Reviewed-by: Richard Levitte <levitte@openssl.org>

This adds a BIO similar to a normal mem BIO but with datagram awareness.
It also has the capability to inject additional packets at arbitrary
locations into the BIO, for testing purposes.

Reviewed-by: Richard Levitte <levitte@openssl.org>

Dump out the records passed over the BIO. Only works for DTLS at the
moment but could easily be extended to TLS.

Reviewed-by: Richard Levitte <levitte@openssl.org>

Reviewed-by: Stephen Henson <steve@openssl.org>

Reviewed-by: Matt Caswell <matt@openssl.org>

Reviewed-by: Matt Caswell <matt@openssl.org>

Reviewed-by: Matt Caswell <matt@openssl.org>

Reviewed-by: Matt Caswell <matt@openssl.org>

@disablables is sorted, but these were just added at the end of
%disabled in commits c2e27310 and 22e3dcb7

.

Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>