Note about GnuTLS incompatibility for updates from 4.50 and earlier.

[exim.git] / src / README.UPDATING

$Cambridge: exim/src/README.UPDATING,v 1.15 2006/10/23 09:14:50 ph10 Exp $

This document contains detailed information about incompatibilities that might
be encountered when upgrading from one release of Exim to another. The
information is in reverse order of release numbers. Mostly these are relatively
small points, and the configuration file is normally upwards compatible, but
there have been two big upheavals...


**************************************************************************
* There was a big reworking of the way mail routing works for release    *
* 4.00. Previously used "directors" were abolished, and all routing is   *
* now done by routers. Policy controls for incoming mail are now done by *
* Access Control Lists instead of separate options. All this means that  *
* pre-4.00 configuration files have to be massively converted. If you    *
* are coming from a 3.xx release, please read the document in the file   *
* doc/Exim4.upgrade, and allow some time to complete the upgrade.        *
*                                                                        *
* There was a big reworking of the way domain/host/net/address lists are *
* handled at release 3.00. If you are coming from a pre-3.00 release, it *
* might be easier to start again from a default configuration. Otherwise *
* you need to read doc/Exim3.upgrade and do a double conversion of your  *
* configuration file.                                                    *
**************************************************************************


The rest of this document contains information about changes in 4.xx releases
that might affect a running system.


Exim version 4.64
-----------------

1. Callouts were setting the name used for EHLO/HELO from $smtp_active_
hostname. This is wrong, because it relates to the incoming message (and
probably the interface on which it is arriving) and not to the outgoing
callout (which could be using a different interface). This has been
changed to use the value of the helo_data option from the smtp transport
instead - this is what is used when a message is actually being sent. If
there is no remote transport (possible with a router that sets up host
addresses), $smtp_active_hostname is used. This change is mentioned here in
case somebody is relying on the use of $smtp_active_hostname.

2. A bug has been fixed that might just possibly be something that is relied on
in some configurations. In expansion items such as ${if >{xxx}{yyy}...} an
empty string (that is {}) was being interpreted as if it was {0} and therefore
treated as the number zero. From release 4.64, such strings cause an error
because a decimal number, possibly followed by K or M, is required (as has
always been documented).

3. There has been a change to the GnuTLS support (ChangeLog/PH/20) to improve
Exim's performance. Unfortunately, this has the side effect of being slightly
non-upwards compatible for versions 4.50 and earlier. If you are upgrading from
one of these earlier versions and you use GnuTLS, you must remove the file
called gnutls-params in Exim's spool directory. If you don't do this, you will
see this error:

  TLS error on connection from ... (DH params import): Base64 decoding error.

Removing the file causes Exim to recompute the relevant encryption parameters
and cache them in the new format that was introduced for release 4.51 (May
2005). If you are upgrading from release 4.51 or later, there should be no
problem.


Exim version 4.63
-----------------

When an SMTP error message is specified in a "message" modifier in an ACL, or
in a :fail: or :defer: message in a redirect router, Exim now checks the start
of the message for an SMTP error code. This consists of three digits followed
by a space, optionally followed by an extended code of the form n.n.n, also
followed by a space. If this is the case and the very first digit is the same
as the default error code, the code from the message is used instead. If the
very first digit is incorrect, a panic error is logged, and the default code is
used. This is an incompatible change, but it is not expected to affect many (if
any) configurations. It is possible to suppress the use of the supplied code in
a redirect router by setting the smtp_error_code option false. In this case,
any SMTP code is quietly ignored.


Exim version 4.61
-----------------

1. The default number of ACL variables of each type has been increased to 20,
and it's possible to compile Exim with more. You can safely upgrade to this
release if you already have messages on the queue with saved ACL variable
values. However, if you downgrade from this release with messages on the queue,
any saved ACL values they may have will be lost.

2. The default value for rfc1413_query_timeout has been changed from 30s to 5s.


Exim version 4.54
-----------------

There was a problem with 4.52/TF/02 in that a "name=" option on control=
submission terminated at the next slash, thereby not allowing for slashes in
the name. This has been changed so that "name=" takes the rest of the string as
its data. It must therefore be the last option.


Version 4.53
------------

If you are using the experimental Domain Keys support, you must upgrade to
at least libdomainkeys 0.67 in order to run this release of Exim.


Version 4.51
------------

1. The format in which GnuTLS parameters are cached (in the file gnutls-params
in the spool directory) has been changed. The new format can also be generated
externally, so it is now possible to update the values from outside Exim. This
has been implemented in an upwards, BUT NOT downwards, compatible manner.
Upgrading should be seamless: when Exim finds that it cannot understand an
existing cache file, it generates new parameters and writes them to the cache
in the new format. If, however, you downgrade from 4.51 to a previous release,
you MUST delete the gnutls-params file in the spool directory, because the
older Exim will not recognize the new format.

2. When doing a callout as part of verifying an address, Exim was not paying
attention to any local part prefix or suffix that was matched by the router
that accepted the address. It now behaves in the same way as it does for
delivery: the affixes are removed from the local part unless
rcpt_include_affixes is set on the transport. If you have a configuration that
uses prefixes or suffixes on addresses that could be used for callouts, and you
want the affixes to be retained, you must make sure that rcpt_include_affixes
is set on the transport.

3. Bounce and delay warning messages no longer contain details of delivery
errors, except for explicit messages (e.g. generated by :fail:) and SMTP
responses from remote hosts.


Version 4.50
------------

The exicyclog script has been updated to use three-digit numbers in rotated log
files if the maximum number to keep is greater than 99. If you are already
keeping more than 99, there will be an incompatible change when you upgrade.
You will probably want to rename your old log files to the new form before
running the new exicyclog.


Version 4.42
------------

RFC 3848 specifies standard names for the "with" phrase in Received: header
lines when AUTH and/or TLS are in use. This is the "received protocol"
field. Exim used to use "asmtp" for authenticated SMTP, without any
indication (in the protocol name) for TLS use. Now it follows the RFC and
uses "esmtpa" if the connection is authenticated, "esmtps" if it is
encrypted, and "esmtpsa" if it is both encrypted and authenticated. These names
appear in log lines as well as in Received: header lines.


Version 4.34
------------

Change 4.31/2 gave problems to data ACLs and local_scan() functions that
expected to see a Received: header. I have changed to yet another scheme. The
Received: header is now generated after the body is received, but before the
ACL or local_scan() is called. After they have run, the timestamp in the
Received: header is updated.

Thus, change (a) of 4.31/2 has been reversed, but change (b) is still true,
which is lucky, since I decided it was a bug fix.


Version 4.33
------------

If an expansion in a condition on a "warn" statement fails because a lookup
defers, the "warn" statement is abandoned, and the next ACL statement is
processed. Previously this caused the whole ACL to be aborted.


Version 4.32
------------

Change 4.31/2 has been reversed, as it proved contentious. Recipient callout
verification now uses <> in the MAIL command by default, as it did before. A
new callout option, "use_sender", has been added to request the other
behaviour.


Version 4.31
------------

1. If you compile Exim to use GnuTLS, it now requires the use of release 1.0.0
   or greater. The interface to the obsolete 0.8.x releases is no longer
   supported. There is one externally visible change: the format for the
   display of Distinguished Names now uses commas as a separator rather than a
   slash. This is to comply with RFC 2253.

2. When a message is received, the Received: header line is now generated when
   reception is complete, instead of at the start of reception. For messages
   that take a long time to come in, this changes the meaning of the timestamp.
   There are several side-effects of this change:

    (a) If a message is rejected by a DATA or non-SMTP ACL, or by local_scan(),
        the logged header lines no longer include the local Received: line,
        because it has not yet been created. If the message is a non-SMTP one,
        and the error is processed by sending a message to the sender, the copy
        of the original message that is returned does not have an added
        Received: line.

    (b) When a filter file is tested using -bf, no additional Received: header
        is added to the test message. After some thought, I decided that this
        is a bug fix.

    The contents of $received_for are not affected by this change. This
    variable still contains the single recipient of a message, copied after
    addresses have been rewritten, but before local_scan() is run.

2. Recipient callout verification, like sender verification, was using <> in
   the MAIL FROM command. This isn't really the right thing, since the actual
   sender may affect whether the remote host accepts the recipient or not. I
   have changed it to use the actual sender in the callout; this means that
   the cache record is now keyed on a recipient/sender pair, not just the
   recipient address. There doesn't seem to be a real danger of callout loops,
   since a callout by the remote host to check the sender would use <>.


Version 4.30
------------

1. I have abolished timeout_DNS as an error that can be detected in retry
   rules, because it has never worked. Despite the fact that it has been
   documented since at least release 1.62, there was no code to support it.
   If you have used it in your retry rules, you will now get a warning message
   to the log and panic log. It is now treated as plain "timeout".

2. After discussion on the mailing list, Exim no longer adds From:, Date:, or
   Message-Id: header lines to messages that do not originate locally, that is,
   messages that have an associated sending host address.

3. When looking up a host name from an IP address, Exim now tries the DNS
   first, and only if that fails does it use gethostbyaddr() (or equivalent).
   This change was made because on some OS, not all the names are given for
   addresses with multiple PTR records via the gethostbyaddr() interface. The
   order of lookup can be changed by setting host_lookup_order.


Version 4.23
------------

1. The new FIXED_NEVER_USERS build-time option creates a list of "never users"
   that cannot be overridden. The default in the distributed EDITME is "root".
   If for some reason you were (against advice) running deliveries as root, you
   will have to ensure that FIXED_NEVER_USERS is not set in your
   Local/Makefile.

2. The ${quote: operator now quotes an empty string, which it did not before.

3. Version 4.23 saves the contents of the ACL variables with the message, so
   that they can be used later. If one of these variables contains a newline,
   there will be a newline character in the spool that will not be interpreted
   correctely by a previous version of Exim. (Exim ignores keyed spool file
   items that it doesn't understand - precisely for this kind of problem - but
   it expects them all to be on one line.)

   So the bottom line is: if you have newlines in your ACL variables, you
   cannot retreat from 4.23.


Version 4.21
------------

1. The idea of the "warn" ACL verb is that it adds a header or writes to the
   log only when "message" or "log_message" are set. However, if one of the
   conditions was an address verification, or a call to a nested ACL, the
   messages generated by the underlying test were being passed through. This
   no longer happens. The underlying message is available in $acl_verify_
   message for both "message" and "log_message" expansions, so it can be
   passed through if needed.

2. The way that the $h_ (and $header_) expansions work has been changed by the
   addition of RFC 2047 decoding. See the main documentation (the NewStuff file
   until release 4.30, then the manual) for full details. Briefly, there are
   now three forms:

     $rh_xxx: and $rheader_xxx: give the original content of the header
       line(s), with no processing at all.

     $bh_xxx: and $bheader_xxx: remove leading and trailing white space, and
       then decode base64 or quoted-printable "words" within the header text,
       but do not do charset translation.

     $h_xxx: and $header_xxx: attempt to translate the $bh_ string to a
       standard character set.

   If you have previously been using $h_ expansions to access the raw
   characters, you should change to $rh_ instead.

3. When Exim creates an RFC 2047 encoded word in a header line, it labels it
   with the default character set from the headers_charset option instead of
   always using iso-8859-1.

4. If TMPDIR is defined in Local/Makefile (default in src/EDITME is
   TMPDIR="/tmp"), Exim checks for the presence of an environment variable
   called TMPDIR, and if it finds it is different, it changes its value.

5. Following a discussion on the list, the rules by which Exim recognises line
   endings on incoming messages have been changed. The -dropcr and drop_cr
   options are now no-ops, retained only for backwards compatibility. The
   following line terminators are recognized: LF CRLF CR. However, special
   processing applies to CR:

   (i)  The sequence CR . CR does *not* terminate an incoming SMTP message,
        nor a local message in the state where . is a terminator.

   (ii) If a bare CR is encountered in a header line, an extra space is added
        after the line terminator so as not to end the header. The reasoning
        behind this is that bare CRs in header lines are most likely either
        to be mistakes, or people trying to play silly games.

6. The code for using daemon_smtp_port, local_interfaces, and the -oX options
   has been reorganized. It is supposed to be backwards compatible, but it is
   mentioned here just in case I've screwed up.



Version 4.20
------------

1. I have tidied and re-organized the code that uses alarm() for imposing time
   limits on various things. It shouldn't affect anything, but if you notice
   processes getting stuck, it may be that I've broken something.

2. The "arguments" log selector now also logs the current working directory
   when Exim is called.

3. An incompatible change has been made to the appendfile transport. This
   affects the case when it is used for file deliveries that are set up by
   .forward and filter files. Previously, any settings of the "file" or
   "directory" options were ignored. It is hoped that, like the address_file
   transport in the default configuration, these options were never in fact set
   on such transports, because they were of no use.

   Now, if either of these options is set, it is used. The path that is passed
   by the router is in $address_file (this is not new), so it can be used as
   part of a longer path, or modified in any other way that expansion permits.

   If neither "file" nor "directory" is set, the behaviour is unchanged.

4. Related to the above: in a filter, if a "save" command specifies a non-
   absolute path, the value of $home/ is pre-pended. This no longer happens if
   $home is unset or is set to an empty string.

5. Multiple file deliveries from a filter or .forward file can never be
   batched; the value of batch_max on the transport is ignored for file
   deliveries. I'm assuming that nobody ever actually set batch_max on the
   address_file transport - it would have had odd effects previously.

6. DESTDIR is the more common variable that ROOT for use when installing
   software under a different root filing system. The Exim install script now
   recognizes DESTDIR first; if it is not set, ROOT is used.

7. If DESTDIR is set when installing Exim, it no longer prepends its value to
   the path of the system aliases file that appears in the default
   configuration (when a default configuration is installed). If an aliases
   file is actually created, its name *does* use the prefix.


Version 4.14
------------

1. The default for the maximum number of unknown SMTP commands that Exim will
accept before dropping a connection has been reduced from 5 to 3. However, you
can now change the value by setting smtp_max_unknown_commands.

2. The ${quote: operator has been changed so that it turns newline and carriage
return characters into \n and \r, respectively.

3. The file names used for maildir messages now include the microsecond time
fraction as well as the time in seconds, to cope with systems where the process
id can be re-used within the same second. The format is now

  <time>.H<microsec>P<pid>.<host>

This should be a compatible change, but is noted here just in case.

4. The rules for creating message ids have changed, to cope with systems where
the process id can be re-used within the same second. The format, however, is
unchanged, so this should not cause any problems, except as noted in the next
item.

5. The maximum value for localhost_number has been reduced from 255 to 16, in
order to implement the new message id rules. For operating systems that have
case-insensitive file systems (Cygwin and Darwin), the limit is 10.

6. verify = header_syntax was allowing unqualified addresses in all cases. Now
it allows them only for locally generated messages and from hosts that match
sender_unqualified_hosts or recipient_unqualified_hosts, respectively.

7. For reasons lost in the mists of time, when a pipe transport was run, the
environment variable MESSAGE_ID was set to the message ID preceded by 'E' (the
form used in Message-ID: header lines). The 'E' has been removed.


Version 4.11
------------

1. The handling of lines in the configuration file has changed. Previously,
macro expansion was applied to logical lines, after continuations had been
joined on. This meant that it could not be used in .include lines, which are
handled as physical rather than logical lines. Macro expansion is now done on
physical lines rather than logical lines. This means there are two
incompatibilities:

  (a) A macro that expands to # to turn a line into a comment now applies only
      to the physical line where it appears. Previously, it would have caused
      any following continuations also to be ignored.

  (b) A macro name can no longer be split over the boundary between a line and
      its continuation. Actually, this is more of a bug fix. :-)

2. The -D command line option must now all be within one command line item.
This makes it possible to use -D to set a macro to the empty string by commands
such as

  exim -DABC  ...
  exim -DABC= ...

Previously, these items would have moved on to the next item on the command
line. To include spaces in a macro definition item, quotes must be used, in
which case you can also have spaces after -D and surrounding the equals. For
example:

  exim '-D ABC = something' ...

3. The way that addresses that redirect to themselves are handled has been
changed, in order to fix an obscure bug. This should not cause any problems
except in the case of wanting to go back from a 4.11 (or later) release to an
earlier release. If there are undelivered messages on the spool that contain
addresses which redirect to themselves, and the redirected addresses have
already been delivered, you might get a duplicate delivery if you revert to an
earlier Exim.

4. The default way of looking up IP addresses for hosts in the manualroute and
queryprogram routers has been changed. If "byname" or "bydns" is explicitly
specified, there is no change, but if no method is specified, Exim now behaves
as follows:

  First, a DNS lookup is done. If this yields anything other than
  HOST_NOT_FOUND, that result is used. Otherwise, Exim goes on to try a call to
  getipnodebyname() (or gethostbyname() on older systems) and the result of the
  lookup is the result of that call.

This change has been made because it has been discovered that on some systems,
if a DNS lookup called via getipnodebyname() times out, HOST_NOT_FOUND is
returned instead of TRY_AGAIN. Thus, it is safest to try a DNS lookup directly
first, and only if that gives a definite "no such host" to try the local
function.

5. In fixing the minor security problem with pid_file_path, I have removed some
backwards-compatible (undocumented) code which was present to ease conversion
from Exim 3. In Exim 4, pid_file_path is a literal; in Exim 3 it was allowed to
contain "%s", which was replaced by the port number for daemons listening on
non-standard ports. In Exim 4, such daemons do not write a pid file. The
backwards compatibility feature was to replace "%s" by nothing if it occurred
in an Exim 4 setting of pid_file_path. The bug was in this code. I have solved
the problem by removing the backwards compatibility feature. Thus, if you still
have "%s" somewhere in a setting of pid_file_path, you should remove it.

6. There has been an extension to lsearch files. The keys in these files may
now be quoted in order to allow for whitespace and colons in them. This means
that if you were previously using keys that began with a doublequote, you will
now have to wrap them with extra quotes and escape the internal quotes. The
possibility that anybody is actually doing this seems extremely remote, but it
is documented just in case.


Version 4.10
------------

The build-time parameter EXIWHAT_KILL_ARG has been renamed EXIWHAT_KILL_SIGNAL
to better reflect its function. The OS-specific files have been updated. Only
if you have explicitly set this in your Makefile (highly unlikely) do you need
to change anything.

****