Fix transport buffer size handling

[exim.git] / doc / doc-txt / GnuTLS-FAQ.txt

Using Exim 4.80+ with GnuTLS
============================

(1) I'm having problems building with GnuTLS 1, why?
(2) What changed?  Why?
(3) I'm seeing:
    "(gnutls_handshake): A TLS packet with unexpected length was received"
    Why?
(4) What's the deal with MD5?  (And SHA-1?)
(5) What happened to gnutls_require_kx / gnutls_require_mac /
    gnutls_require_protocols?
(6) What's the deal with tls_dh_max_bits?  What's DH?
(7) What's a Priority String?
(8) How do I use tls_require_ciphers?
(9) How do I test STARTTLS support?



(1): I'm having problems building with GnuTLS 1, why?
-----------------------------------------------------

GnuTLS's library interface has changed and Exim uses the more current
interface.  Since GnuTLS is security critical code, you should probably update
to a supported release.

If updating GnuTLS is not an option, then build Exim against OpenSSL instead.

If neither is an option, then you might build Exim with the rule
"SUPPORT_TLS=yes" commented out in "Local/Makefile", so that your Exim build
no longer has TLS support.

If you need to keep TLS support, and you can't use OpenSSL, then you'll have
to update the GnuTLS you have installed.  Sorry.

We've tested the build of Exim back as far as GnuTLS 2.8.x; most development
work is done with 2.12 and tested on 2.10 and 3.x.

If you have to pick a version to upgrade to, use GnuTLS 3.x if available.  The
GnuTLS developers took advantage of the version bump to add an error code
return value which makes debugging some problems a lot easier.



(2): What changed?  Why?
------------------------

The GnuTLS provider integration in Exim was overhauled, rewritten but with
some copy/paste, because building Exim against more current releases of GnuTLS
was issuing deprecation warnings from the compiler.

When a library provider marks up the include files so that some function calls
will cause the compiler/linker to emit deprecation warnings, it's time to pay
serious attention.  A future release might not work at all.  Using the new
APIs may mean that Exim will *stop* working with older releases of GnuTLS.
The GnuTLS support in Exim was overhauled in Exim 4.80.  In prior releases,
Exim hard-coded a lot of algorithms and constrained what could happen.  In
Exim 4.79, we added to the hard-coded list just enough to let TLSv1.1 and
TLSv1.2 be negotiated, but not actually support the mandatory algorithms of
those protocol versions.  When Exim's GnuTLS integration was originally
written, there was no other choice than to make Exim responsible for a lot of
this.  In the meantime, GnuTLS has improved.

With the rewrite, we started using the current API and leaving a lot more
responsibility for TLS decisions to the library.

The GnuTLS developers added "priority strings" (see Q7), which provide an
interface exposed to the configuration file for a lot of the tuning.

The GnuTLS policy is to no longer support MD5 in certificates.  Exim had
previously been immune to this policy, but no longer.  See Q4.



(3): I'm seeing "A TLS packet with unexpected length was received".  Why?
-------------------------------------------------------------------------

The most likely reason is that the client dropped the connection during
handshake, because their library disliked some aspect of the negotiation.

In GnuTLS 2, an EOF on the connection is reported with an error code for
packets being too large, and the above is the string returned by the library
for that error code.  In GnuTLS 3, there's a specific error code for EOF and
the diagnostic will be less confusing.

Most likely cause is an MD5 hash used in a certificate.  See Q4 below.
Alternatively, the client dislikes the size of the Diffie-Hellman prime
offered by the server; if lowering the value of the "tls_dh_max_bits" Exim
option fixes the problem, this was the cause.  See Q6.



(4): What's the deal with MD5?  (And SHA-1?)
--------------------------------------------

MD5 is a hash algorithm.  Hash algorithms are used to reduce a lot of data
down to a fairly short value, which is supposed to be extremely hard to
manipulate to get a value of someone's choosing.  Signatures, used to attest
to identity or integrity, rely upon this manipulation being effectively
impossible, because the signature is the result of math upon the hash result.
Without hash algorithms, signatures would be longer than the text being
signed.

MD5 was once very popular.  It still is far too popular.  Real world attacks
have been proven possible against MD5.  Including an attack against PKI
(Public Key Infrastructure) certificates used for SSL/TLS.  In that attack,
the attackers got a certificate for one identity but were able to then publish
a certificate with the same signature but a different identity.  This
undermines the whole purpose of having certificates.

So GnuTLS stopped trusting any certificate with an MD5-based hash used in it.
The world has been hurriedly moving away from MD5 in certificates for a while.
If you still have such a certificate, you should move too.

If you paid someone for your certificate, they should be willing to reissue
the certificate with a different algorithm, for no extra money.  If they try
to charge money to replace their defective product, buy from someone else
instead.  Part of the reason for paying money on a recurring basis is to cover
the ongoing costs of proving a trust relationship, such as providing
revocation protocols.  This is just another of those ongoing costs you have
already paid for.

The same has happened to SHA-1: there are real-world collision attacks against
SHA-1, so SHA-1 is mostly defunct in certificates.  GnuTLS no longer supports
its use in TLS certificates.



(5): ... gnutls_require_kx / gnutls_require_mac / gnutls_require_protocols?
---------------------------------------------------------------------------

These Exim options were used to provide fine-grained control over the TLS
negotiation performed by GnuTLS.  They required explicit protocol knowledge
from Exim, which vastly limited what GnuTLS could do and involved the Exim
maintainers in decisions which aren't part of their professional areas of
expertise.  The need for Exim to be able to do this went away when GnuTLS
introduced Priority Strings (see Q7).

If you were using these options before, then you're already an expert user and
should be able to easily craft a priority string to accomplish your goals.
Set the Exim "tls_require_ciphers" value accordingly.  There is a main section
option of this name, used for Exim receiving inbound connections, and an SMTP
driver transport option of this name, used for Exim establishing outbound
connections.



(6): What's the deal with tls_dh_max_bits?  What's DH?
------------------------------------------------------

You can avoid all of the tls_dh_max_bits issues if you leave "tls_dhparam"
unset, so that you get one of the standard built-in primes used for DH.


DH, Diffie-Hellman (or Diffie-Hellman-Merkle, or something naming Williamson)
is the common name for a way for two parties to a communication stream to
exchange some private random data so that both end up with a shared secret
which no eavesdropper can get.  It does not provide for proof of the identity
of either party, so on its own is subject to man-in-the-middle attacks, but is
often combined with systems which do provide such proof, improving them by
separating the session key (the shared secret) from the long-term identity,
and so protecting past communications from a break of the long-term identity.

To do this, the server sends to the client a very large prime number; this is
in the clear, an attacker can see it.  This is not a problem; it's so not a
problem, that there are standard named primes which applications can use, and
which Exim now supports.

The size of the prime number affects how difficult it is to break apart the
shared secret and decrypt the data.  As time passes, the size required to
provide protection against an adversary climbs: computers get more powerful,
mathematical advances are made, and so on.

Estimates of the size needed are published as recommendations by various
groups; a good summary of sizes currently recommended, for various
cryptographic primitives, is available at:

  http://www.keylength.com/en/3/

The GnuTLS folks think the ECRYPT II advice is good.  They know far more of
such matters than the Exim folks, we just say "er, what they said".

One of the new pieces of the GnuTLS API is a means for an application to ask
it for guidance and advice on how large some numbers should be.  This is not
entirely internal to GnuTLS, since generating the numbers is slow, an
application might want to use a standard prime, etc.  So, in an attempt to get
away from being involved in cryptographic policy, and to get rid of a
hard-coded "1024" in Exim's source-code, we switched to asking GnuTLS how many
bits should be in the prime number generated for use for Diffie-Hellman.  We
then give this number straight back to GnuTLS when generating a DH prime.
We can ask for various sizes, and did not expose this to the administrator but
instead just asked for "NORMAL" protection.
Literally:

 dh_bits = gnutls_sec_param_to_pk_bits(GNUTLS_PK_DH, GNUTLS_SEC_PARAM_NORMAL);

This API is only available as of GnuTLS 2.12.  Prior to that release, we stuck
with the old value, for compatibility, so "1024" is still hard-coded.
Reviewing the page above, you'll see that this is described as "Short-term
protection against medium organizations, medium-term protection against small
organizations."

So if you are using an old release of GnuTLS, you can either add to
Local/Makefile a different value of "EXIM_SERVER_DH_BITS_PRE2_12" or accept
that your protection might not be adequate to your needs.  We advise updating
to a more current GnuTLS release and rebuilding Exim against that.

Unfortunately, some TLS libraries have the client side bound how large a DH
prime they will accept from the server.  The larger the number, the more
computation required to work with it and the slower that things get.  So they
pick what they believe to be reasonable upper bounds, and then typically
forget about it for several years.

Worse, in TLS the DH negotiation happens after a ciphersuite has been chosen,
so if the client dislikes the value then a different ciphersuite avoiding DH
can not be negotiated!  The client typically drops the connection, resulting
in errors to the user and errors in the Exim logs.  With GnuTLS 3, you'll see
the EOF (End-Of-File) error message in Exim's logs, reported as being part of
"gnutls_handshake", but with GnuTLS 2 you'll see a log message about a packet
with an unexpected size.  Unless the client software is written intelligently
enough to be able to adapt and reconnect forbidding DH, the client will never
be able to negotiate TLS.

This time around, we discovered that the NSS library used by various Mozilla
products, Chrome, etc, and most particularly by the Thunderbird mail client,
has the lowest cap.  In fact, prior to recent updates, their upper limit was
lower than the value returned by GnuTLS for "NORMAL".  The most recent NSS
library release raises this, but the most recent Thunderbird release still has
the old limit.

So Exim had to get involved in cryptography policy decisions again.  We added
the "tls_dh_max_bits" global option, to set a number used in both OpenSSL and
GnuTLS bindings for Exim.  In GnuTLS, it clamps the value returned by
gnutls_sec_param_to_pk_bits(), so that if the returned value is larger than
tls_dh_max_bits then tls_dh_max_bits would be used instead.

Our policy decision was to default the value of tls_dh_max_bits to the maximum
supported in the most recent Thunderbird release, and to make this an
administrator-available option so that administrators can choose to trade off
security versus compatibility by raising it.

A future release of Exim may even let the administrator tell GnuTLS to ask for
more or less than "NORMAL".

To add to the fun, the size of the prime returned by GnuTLS when we call
gnutls_dh_params_generate2() is not limited to be the requested size.  GnuTLS
has a tendency to overshoot.  2237 bit primes are common when 2236 is
requested, and higher still have been observed.  Further, there is no API to
ask how large the prime bundled up inside the parameter is; the most we can do
is ask how large the DH prime used in an active TLS session is.  Since we're
not able to use GnuTLS API calls (and exporting to PKCS3 and then calling
OpenSSL routines would be undiplomatic, plus add a library dependency), we're
left with no way to actually know the size of the freshly generated DH prime.

Thus we check if the the value returned is at least 10 more than the minimum
we'll accept as a client (EXIM_CLIENT_DH_MIN_BITS, see below, defaults to
1024) and if it is, we subtract 10.  Then we reluctantly deploy a strategy
called "hope".  This is not guaranteed to be successful; in the first code
pass on this logic, we subtracted 3, asked for 2233 bits and got 2240 in the
first test.

If you see Thunderbird clients still failing, then as a user who can see into
Exim's spool directory, run:

$ openssl dhparam -noout -text -in /path/to/spool/gnutls-params-2236 | head

Ideally, the first line will read "PKCS#3 DH Parameters: (2236 bit)".  If the
count is more than 2236, then remove the file and let Exim regenerate it, or
generate one yourself and move it into place.  Ideally use "openssl dhparam"
to generate it, and then wait a very long time; at least this way, the size
will be correct.

The use of "hope" as a strategy was felt to be unacceptable as a default, so
late in the RC series for 4.80, the whole issue was side-stepped.  The primes
used for DH are publicly revealed; moreover, there are selection criteria for
what makes a "good" DH prime.  As it happens, there are *standard* primes
which can be used, and are specified to be used for certain protocols.  So
these primes were built into Exim, and by default exim now uses a 2048 bit
prime from section 2.2 of RFC 5114.


A TLS client does not get to choose the DH prime used, but can choose a
minimum acceptable value.  For Exim, this is a compile-time constant called
"EXIM_CLIENT_DH_MIN_BITS" of 1024, which can be overruled in "Local/Makefile".



(7): What's a Priority String?
------------------------------

A priority string is a way for a user of GnuTLS to tell GnuTLS how it should
make decisions about what to do in TLS; it includes which algorithms to make
available for various roles, what compatibility trade-offs to make, which
features to enable or disable.

It is exposed to the Mail Administrator in Exim's configuration file as the
"tls_require_ciphers" option, which exists as a main section option for use in
Exim as a server, accepting connections, and as an option on Transports using
the SMTP driver, for use in Exim as a client.  The main section option is
*not* the default for the transport option, they are entirely independent.
For both, the default value used by Exim is the string "NORMAL".  (This is not
the same NORMAL as for DH prime bit size selection in Q6, but a different
NORMAL.)  See Q8.

The current documentation, for the most recent release of GnuTLS, is available
online at:

  http://www.gnutls.org/manual/html_node/Priority-Strings.html

Beware that if you are not using the most recent GnuTLS release then this
documentation will be wrong for you!  You should find the "info" documentation
which came with GnuTLS to review the available options.  It's under "The TLS
Handshake Protocol".

$ pinfo --node="Priority Strings" gnutls

(This author is unable to persuade the "info" command-line tool to jump
straight to the required node, but "pinfo" works.)

To trade off some security for more compatibility, you might set a value of
"NORMAL:%COMPAT".  See the documentation for more, including lowering security
even further for more security, forcing clients to use the server's protocol
suite, and ways to force selection of particular algorithms.



(8): How do I use tls_require_ciphers?
--------------------------------------

This is the name of two options in Exim.  One is a main section option, used
by Exim as a server when a client initiates SSL/TLS negotiation, the other is
an option on transports which use "driver = smtp", used when Exim initiates
SSL/TLS as a client talking to a remote server.

The option is expanded and so can take advantage of any variables which have
been set.  This includes the IP address of the remote side, the port upon
which a connection was accepted (when a server), and more.  Currently it does
not have access to $tls_sni, whether as a client or as a server.

This example, for the main section's option, will let the library defaults be
permitted on the MX port, where there's probably no identity verification
anyway, and lowers security further by increasing compatibility; but this ups
the ante on the submission ports where the administrator might have some
influence on the choice of clients used:

tls_require_ciphers = ${if =={$received_port}{25}\
                           {NORMAL:%COMPAT}\
                           {SECURE128}}

Note that during Exim start-up, when this option is sanity-checked, there will
be no value of $received_port.  In the above example, the checked value will
thus be "SECURE128".  Be careful to ensure that it always expands safely.



(9): How do I test STARTTLS support?
------------------------------------

The best command-line client for debugging specifically SSL/TLS which this
author has encountered is part of the GnuTLS suite, and is called
"gnutls-cli".  It's best because it's the only interactive tool which lets the
user start TLS handshake exactly when they wish, so can choose to use the
STARTTLS command.

$ gnutls-cli --starttls --crlf --port 587 mail.example.org

After EHLO, to see the capabilities, enter STARTTLS, wait for the response,
then send EOF.  Typically that's done by typing Ctrl-D at the start of a line.
The "gnutls-cli" tool will take over, set up TLS (or fail) and by the time it
returns to await more user input, you're using a secure connection and should
type your second EHLO.

The "--x509cafile" option may be helpful for checking certificates and
"--priority" to pass a priority string to the client tool for configuring it.

The --crlf is for strict protocol correctness, but Exim doesn't really need
it, so "gnutls-cli -s -p 587 mail.example.org" is shorter.


For debugging SMTP as a whole, we recommend swaks, "Swiss Army Knife SMTP", by
John Jetmore (one of the Exim Maintainers).  This has some TLS tuning options;
it can be found at:

  http://www.jetmore.org/john/code/swaks/


For OpenSSL, the "openssl s_client" command helps; you can either set up Exim
with a listening port which is SSL-on-connect or tell s_client to use
STARTTLS.

For the former, use the "tls_on_connect_ports" option and the
"daemon_smtp_ports" option.  Most clients for SSL-on-connect use the port
which was briefly registered with IANA for this purpose, 465.  So you would
set something like:

  daemon_smtp_ports = 25 : 465 : 587
  tls_on_connect_ports = 465

To use s_client with STARTTLS support, use "-starttls smtp" on the
command-line.  Beware that older versions of OpenSSL did not wait for the SMTP
banner before sending EHLO, which will fall afoul of the protocol
synchronisation checks in Exim (used to trip up pump-and-dump spammers); also
you will not get control of the session until TLS is established.  That said,
this tool provides more tuning hooks for adjusting how TLS will be set up than
most.

*BEWARE* that by default, s_client will take any line starting with a capital
letter "R" to be a request to initiate TLS renegotiation with the server and
the line will not be sent.  This may trip up "RCPT TO:<someone@example.org>"
lines in SMTP.  SMTP is not case-sensitive, so type "rcpt to" instead.
Alternatively, invoke s_client with the "-ign_eof" option to disable this
R-filtering and a few other features.


# END OF FAQ