X-Git-Url: https://git.exim.org/users/jgh/exim.git/blobdiff_plain/5fe762f68d697d0958c02da6ba18a2256eaf134d..54a4fbd526353c5ec1ce25ebe9c3b48f05b1e5be:/doc/doc-txt/ChangeLog

diff --git a/doc/doc-txt/ChangeLog b/doc/doc-txt/ChangeLog
index ef68f01af..77e858762 100644
--- a/doc/doc-txt/ChangeLog
+++ b/doc/doc-txt/ChangeLog
@@ -1,17 +1,562 @@
-$Cambridge: exim/doc/doc-txt/ChangeLog,v 1.2 2004/10/11 09:49:43 ph10 Exp $
+$Cambridge: exim/doc/doc-txt/ChangeLog,v 1.78 2005/02/15 09:31:13 ph10 Exp $
 
 Change log file for Exim from version 4.21
 -------------------------------------------
 
 
-Exim version 4.44
+A note about Exim versions 4.44 and 4.50
+----------------------------------------
+
+Exim 4.50 was meant to be the next release after 4.43. It contains a lot of
+changes of various kinds. As a consequence, a big documentation update was
+needed. This delayed the release for rather longer than seemed good, especially
+in the light of a couple of (minor) security issues. Therefore, the changes
+that fixed bugs were backported into 4.43, to create a 4.44 maintenance
+release. So 4.44 and 4.50 are in effect two different branches that both start
+from 4.43.
+
+I have left the 4.50 change log unchanged; it contains all the changes since
+4.43. The change log for 4.44 is below; many of its items are identical to
+those for 4.50. This seems to be the most sensible way to preserve the
+historical information.
+
+
+Exim version 4.50
 -----------------
 
  1. Minor wording change to the doc/README.SIEVE file.
 
- 2. Change 4.43/35 introduced a bug: if quota_filecount was set, the 
+ 2. Change 4.43/35 introduced a bug: if quota_filecount was set, the
     computation of the current number of files was incorrect.
 
+ 3. Closing a stable door: arrange to panic-die if setitimer() ever fails. The
+    bug fixed in 4.43/37 would have been diagnosed quickly if this had been in
+    place.
+
+ 4. Give more explanation in the error message when the command for a transport
+    filter fails to execute.
+
+ 5. There are several places where Exim runs a non-Exim command in a
+    subprocess. The SIGUSR1 signal should be disabled for these processes. This
+    was being done only for the command run by the queryprogram router. It is
+    now done for all such subprocesses. The other cases are: ${run, transport
+    filters, and the commands run by the lmtp and pipe transports.
+
+ 6. Added CONFIGURE_GROUP build-time option.
+
+ 7. Some older OS have a limit of 256 on the maximum number of file
+    descriptors. Exim was using setrlimit() to set 1000 as a large value
+    unlikely to be exceeded. Change 4.43/17 caused a lot of logging on these
+    systems. I've change it so that if it can't get 1000, it tries for 256.
+
+ 8. "control=submission" was allowed, but had no effect, in a DATA ACL. This
+    was an oversight, and furthermore, ever since the addition of extra
+    controls (e.g. 4.43/32), the checks on when to allow different forms of
+    "control" were broken. There should now be diagnostics for all cases when a
+    control that does not make sense is encountered.
+
+ 9. Added the /retain_sender option to "control=submission".
+
+10. $recipients is now available in the predata ACL (oversight).
+
+11. Tidy the search cache before the fork to do a delivery from a message
+    received from the command line. Otherwise the child will trigger a lookup
+    failure and thereby defer the delivery if it tries to use (for example) a
+    cached ldap connection that the parent has called unbind on.
+
+12. If verify=recipient was followed by verify=sender in a RCPT ACL, the value
+    of $address_data from the recipient verification was clobbered by the
+    sender verification.
+
+13. The value of address_data from a sender verification is now available in
+    $sender_address_data in subsequent conditions in the ACL statement.
+
+14. Added forbid_sieve_filter and forbid_exim_filter to the redirect router.
+
+15. Added a new option "connect=<time>" to callout options, to set a different
+    connection timeout.
+
+16. If FIXED_NEVER_USERS was defined, but empty, Exim was assuming the uid 0
+    was its contents. (It was OK if the option was not defined at all.)
+
+17. A "Completed" log line is now written for messages that are removed from
+    the spool by the -Mrm option.
+
+18. New variables $sender_verify_failure and $recipient_verify_failure contain
+    information about exactly what failed.
+
+19. Added -dd to debug only the daemon process.
+
+20. Incorporated Michael Haardt's patch to ldap.c for improving the way it
+    handles timeouts, both on the server side and network timeouts. Renamed the
+    CONNECT parameter as NETTIMEOUT (but kept the old name for compatibility).
+
+21. The rare case of EHLO->STARTTLS->HELO was setting the protocol to "smtp".
+    It is now set to "smtps".
+
+22. $host_address is now set to the target address during the checking of
+    ignore_target_hosts.
+
+23. When checking ignore_target_hosts for an ipliteral router, no host name was
+    being passed; this would have caused $sender_host_name to have been used if
+    matching the list had actually called for a host name (not very likely,
+    since this list is usually IP addresses). A host name is now passed as
+    "[x.x.x.x]".
+
+24. Changed the calls that set up the SIGCHLD handler in the daemon to use the
+    code that specifies a non-restarting handler (typically sigaction() in
+    modern systems) in an attempt to fix a rare and obscure crash bug.
+
+25. Narrowed the window for a race in the daemon that could cause it to ignore
+    SIGCHLD signals. This is not a major problem, because they are used only to
+    wake it up if nothing else does.
+
+26. A malformed maildirsize file could cause Exim to calculate negative values
+    for the mailbox size or file count. Odd effects could occur as a result.
+    The maildirsize information is now recalculated if the size or filecount
+    end up negative.
+
+27. Added HAVE_SYS_STATVFS_H to the os.h file for Linux, as it has had this
+    support for a long time. Removed HAVE_SYS_VFS_H.
+
+28. Installed the latest version of exipick from John Jetmore.
+
+29. In an address list, if the pattern was not a regular expression, an empty
+    subject address (from a bounce message) matched only if the pattern was an
+    empty string. Non-empty patterns were not even tested. This was the wrong
+    because it is perfectly reasonable to use an empty address as part of a
+    database query. An empty address is now tested by patterns that are
+    lookups. However, all the other forms of pattern expect the subject to
+    contain a local part and a domain, and therefore, for them, an empty
+    address still always fails if the pattern is not itself empty.
+
+30. Exim went into a mad DNS loop when attempting to do a callout where the
+    host was specified on an smtp transport, and looking it up yielded more
+    than one IP address.
+
+31. Re-factored the code for checking spool and log partition space into a
+    function that finds that data and another that does the check. The former
+    is then used to implement four new variables: $spool_space, $log_space,
+    $spool_inodes, and $log_inodes.
+
+32. The RFC2047 encoding function was originally intended for short strings
+    such as real names; it was not keeping to the 75-character limit for
+    encoded words that the RFC imposes. It now respects the limit, and
+    generates multiple encoded words if necessary. To be on the safe side, I
+    have increased the buffer size for the ${rfc2047: expansion operator from
+    1024 to 2048 bytes.
+
+33. It is now permitted to omit both strings after an "if" condition; if the
+    condition is true, the result is "true". As before, when the second string
+    is omitted, a false condition yields an empty string. This makes it less
+    cumbersome to write custom ACL and router conditions.
+
+34. Failure to deliver a bounce message always caused it to be frozen, even if
+    there was an errors_to setting on the router. The errors_to setting is now
+    respected.
+
+35. If an IPv6 address is given for -bh or -bhc, it is now converted to the
+    canonical form (fully expanded) before being placed in
+    $sender_host_address.
+
+36. The table in the code that translates DNS record types into text (T_A to
+    "A" for instance) was missing entries for NS and CNAME. It is just possible
+    that this could have caused confusion if both these types were looked up
+    for the same domain, because the text type is used as part of Exim's
+    per-process caching. But the chance of anyone hitting this buglet seems
+    very small.
+
+37. The dnsdb lookup has been extended in a number of ways.
+
+    (1) There is a new type, "zns", which walks up the domain tree until it
+        finds some nameserver records. It should be used with care.
+
+    (2) There is a new type, "mxh", which is like "mx" except that it returns
+        just the host names, not the priorities.
+
+    (3) It is now possible to give a list of domains (or IP addresses) to be
+        looked up. The behaviour when one of the lookups defers can be
+        controlled by a keyword.
+
+    (4) It is now possible to specify the separator character for use when
+        multiple records are returned.
+
+38. The dnslists ACL condition has been extended: it is now possible to supply
+    a list of IP addresses and/or domains to be looked up in a particular DNS
+    domain.
+
+39. Added log_selector=+queue_time_overall.
+
+40. When running the queue in the test harness, wait just a tad after forking a
+    delivery process, to get repeatability of debugging output.
+
+41. Include certificate and key file names in error message when GnuTLS fails
+    to set them up, because the GnuTLS error message doesn't include the name
+    of the failing file when there is a problem reading it.
+
+42. Allow both -bf and -bF in the same test run.
+
+43. Did the same fix as 41 above for OpenSSL, which had the same infelicity.
+
+44. The "Exiscan patch" is now merged into the mainline Exim source.
+
+45. Sometimes the final signoff response after QUIT could fail to get
+    transmitted in the non-TLS case. Testing !tls_active instead of tls_active
+    < 0 before doing a fflush(). This bug looks as though it goes back to the
+    introduction of TLS in release 3.20, but "sometimes" must have been rare
+    because the tests only now provoked it.
+
+46. Reset the locale to "C" after calling embedded Perl, in case it was changed
+    (this can affect the format of dates).
+
+47. exim_tidydb, when checking for the continued existence of a message for
+    which it has found a message-specific retry record, was not finding
+    messages that were in split spool directories. Consequently, it was
+    deleting retry records that should have stayed in existence.
+
+48. Steve fixed some bugs in eximstats.
+
+49. The SPA authentication driver was not abandoning authentication and moving
+    on to the next authenticator when an expansion was forced to fail,
+    contradicting the general specification for all authenticators. Instead it
+    was generating a temporary error. It now behaves as specified.
+
+50. The default ordering of permitted cipher suites for GnuTLS was pessimal
+    (the order specifies the preference for clients). The order is now AES256,
+    AES128, 3DES, ARCFOUR128.
+
+51. Small patch to Sieve code - explicitly set From: when generating an
+    autoreply.
+
+52. Exim crashed if a remote delivery caused a very long error message to be
+    recorded - for instance if somebody sent an entire SpamAssassin report back
+    as a large number of 550 error lines. This bug was coincidentally fixed by
+    increasing the size of one of Exim's internal buffers (big_buffer) that
+    happened as part of the Exiscan merge. However, to be on the safe side, I
+    have made the code more robust (and fixed the comments that describe what
+    is going on).
+
+53. Now that there can be additional text after "Completed" in log lines (if
+    the queue_time_overall log selector is set), a one-byte patch to exigrep
+    was needed to allow it to recognize "Completed" as not the last thing in
+    the line.
+
+54. The LDAP lookup was not handling a return of LDAP_RES_SEARCH_REFERENCE. A
+    patch that reportedly fixes this has been added. I am not expert enough to
+    create a test for it. This is what the patch creator wrote:
+
+      "I found a little strange behaviour of ldap code when working with
+      Windows 2003 AD Domain, where users was placed in more than one
+      Organization Units. When I tried to give exim partial DN, the exit code
+      of ldap_search was unknown to exim because of LDAP_RES_SEARCH_REFERENCE.
+      But simultaneously result of request was absolutely normal ldap result,
+      so I produce this patch..."
+
+    Later: it seems that not all versions of LDAP support LDAP_RES_SEARCH_
+    REFERENCE, so I have modified the code to exclude the patch when that macro
+    is not defined.
+
+55. Some experimental protocols are using DNS PTR records for new purposes. The
+    keys for these records are domain names, not reversed IP addresses. The
+    dnsdb PTR lookup now tests whether its key is an IP address. If not, it
+    leaves it alone. Component reversal etc. now happens only for IP addresses.
+    CAN-2005-0021
+
+56. Improve error message when ldap_search() fails in OpenLDAP or Solaris LDAP.
+
+57. Double the size of the debug message buffer (to 2048) so that more of very
+    long debug lines gets shown.
+
+58. The exicyclog utility now does better if the number of log files to keep
+    exceeds 99. In this case, it numbers them 001, 002 ... instead of 01, 02...
+
+59. Two changes related to the smtp_active_hostname option:
+
+      (1) $smtp_active_hostname is now available as a variable.
+      (2) The default for smtp_banner uses $smtp_active_hostname instead
+          of $primary_hostname.
+
+60. The host_aton() function is supposed to be passed a string that is known
+    to be a valid IP address. However, in the case of IPv6 addresses, it was
+    not checking this. This is a hostage to fortune. Exim now panics and dies
+    if the condition is not met. A case was found where this could be provoked
+    from a dnsdb PTR lookup with an IPv6 address that had more than 8
+    components; fortuitously, this particular loophole had already been fixed
+    by change 4.50/55 above.
+
+    If there are any other similar loopholes, the new check in host_aton()
+    itself should stop them being exploited. The report I received stated that
+    data on the command line could provoke the exploit when Exim was running as
+    exim, but did not say which command line option was involved. All I could
+    find was the use of -be with a bad dnsdb PTR lookup, and in that case it is
+    running as the user.
+    CAN-2005-0021
+
+61. There was a buffer overflow vulnerability in the SPA authentication code
+    (which came originally from the Samba project). I have added a test to the
+    spa_base64_to_bits() function which I hope fixes it.
+    CAN-2005-0022
+
+62. Configuration update for GNU/Hurd and variations. Updated Makefile-GNU and
+    os.h-GNU, and added configuration files for GNUkFreeBSD and GNUkNetBSD.
+
+63. The daemon start-up calls getloadavg() while still root for those OS that
+    need the first call to be done as root, but it missed one case: when
+    deliver_queue_load_max is set with deliver_drop_privilege. This is
+    necessary for the benefit of the queue runner, because there is no re-exec
+    when deliver_drop_privilege is set.
+
+64. A call to exiwhat cut short delays set up by "delay" modifiers in ACLs.
+    This has been fixed.
+
+65. Caching of lookup data for "hosts =" ACL conditions, when a named host list
+    was in use, was not putting the data itself into the right store pool;
+    consequently, it could be overwritten for a subsequent message in the same
+    SMTP connection. (Fix 4.40/11 dealt with the non-cache case, but overlooked
+    the caching.)
+
+66. Added hosts_max_try_hardlimit to the smtp transport, default 50.
+
+67. The string_is_ip_address() function returns 0, 4, or 6, for "no an IP
+    address", "IPv4 address", and "IPv6 address", respectively. Some calls of
+    the function were treating the return as a boolean value, which happened to
+    work because 0=false and not-0=true, but is not correct code.
+
+68. The host_aton() function was not handling scoped IPv6 addresses (those
+    with, for example, "%eth0" on the end) correctly.
+
+69. Fixed some compiler warnings in acl.c for the bitmaps specified with
+    negated items (that is, ~something) in unsigned ints. Some compilers
+    apparently mutter when there is no cast.
+
+70. If an address verification called from an ACL failed, and did not produce a
+    user-specific message (i.e. there was only a "system" message), nothing was
+    put in $acl_verify_message. In this situation, it now puts the system
+    message there.
+
+71. Change 4.23/11 added synchronization checking at the start of an SMTP
+    session; change 4.31/43 added the unwanted input to the log line - except
+    that it did not do this in the start of session case. It now does.
+
+72. After a timeout in a callout SMTP session, Exim still sent a QUIT command.
+    This is wrong and can cause the other end to generate a synchronization
+    error if it is another Exim or anything else that does the synchronization
+    check. A QUIT command is no longer sent after a timeout.
+
+73. $host_lookup_deferred has been added, to make it easier to detect DEFERs
+    during host lookups.
+
+74. The defer_ok option of callout verification was not working if it was used
+    when verifying addresses in header lines, that is, for this case:
+
+      verify = header_sender/callout=defer_ok
+
+75. A backgrounded daemon closed stdin/stdout/stderr on entry; this meant that
+    those file descriptors could be used for SMTP connections. If anything
+    wrote to stderr (the example that came up was "warn" in embedded Perl), it
+    could be sent to the SMTP client, causing chaos. The daemon now opens
+    stdin, stdout, and stderr to /dev/null when it puts itself into the
+    background.
+
+76. Arrange for output from Perl's "warn" command to be written to Exim's main
+    log by default. The user can override this with suitable Perl magic.
+
+77. The use of log_message on a "discard" ACL verb, which is supposed to add to
+    the log message when discard triggers, was not working for the DATA ACL or
+    for the non-SMTP ACL.
+
+78. Error message wording change in sieve.c.
+
+
+----------------------------------------------------
+See the note above about the 4.44 and 4.50 releases.
+----------------------------------------------------
+
+
+Exim version 4.44
+-----------------
+
+ 1. Change 4.43/35 introduced a bug that caused file counts to be
+    incorrectly computed when quota_filecount was set in an appendfile
+    transport
+
+ 2. Closing a stable door: arrange to panic-die if setitimer() ever fails. The
+    bug fixed in 4.43/37 would have been diagnosed quickly if this had been in
+    place.
+
+ 3. Give more explanation in the error message when the command for a transport
+    filter fails to execute.
+
+ 4. There are several places where Exim runs a non-Exim command in a
+    subprocess. The SIGUSR1 signal should be disabled for these processes. This
+    was being done only for the command run by the queryprogram router. It is
+    now done for all such subprocesses. The other cases are: ${run, transport
+    filters, and the commands run by the lmtp and pipe transports.
+
+ 5. Some older OS have a limit of 256 on the maximum number of file
+    descriptors. Exim was using setrlimit() to set 1000 as a large value
+    unlikely to be exceeded. Change 4.43/17 caused a lot of logging on these
+    systems. I've change it so that if it can't get 1000, it tries for 256.
+
+ 6. "control=submission" was allowed, but had no effect, in a DATA ACL. This
+    was an oversight, and furthermore, ever since the addition of extra
+    controls (e.g. 4.43/32), the checks on when to allow different forms of
+    "control" were broken. There should now be diagnostics for all cases when a
+    control that does not make sense is encountered.
+
+ 7. $recipients is now available in the predata ACL (oversight).
+
+ 8. Tidy the search cache before the fork to do a delivery from a message
+    received from the command line. Otherwise the child will trigger a lookup
+    failure and thereby defer the delivery if it tries to use (for example) a
+    cached ldap connection that the parent has called unbind on.
+
+ 9. If verify=recipient was followed by verify=sender in a RCPT ACL, the value
+    of $address_data from the recipient verification was clobbered by the
+    sender verification.
+
+10. If FIXED_NEVER_USERS was defined, but empty, Exim was assuming the uid 0
+    was its contents. (It was OK if the option was not defined at all.)
+
+11. A "Completed" log line is now written for messages that are removed from
+    the spool by the -Mrm option.
+
+12. $host_address is now set to the target address during the checking of
+    ignore_target_hosts.
+
+13. When checking ignore_target_hosts for an ipliteral router, no host name was
+    being passed; this would have caused $sender_host_name to have been used if
+    matching the list had actually called for a host name (not very likely,
+    since this list is usually IP addresses). A host name is now passed as
+    "[x.x.x.x]".
+
+14. Changed the calls that set up the SIGCHLD handler in the daemon to use the
+    code that specifies a non-restarting handler (typically sigaction() in
+    modern systems) in an attempt to fix a rare and obscure crash bug.
+
+15. Narrowed the window for a race in the daemon that could cause it to ignore
+    SIGCHLD signals. This is not a major problem, because they are used only to
+    wake it up if nothing else does.
+
+16. A malformed maildirsize file could cause Exim to calculate negative values
+    for the mailbox size or file count. Odd effects could occur as a result.
+    The maildirsize information is now recalculated if the size or filecount
+    end up negative.
+
+17. Added HAVE_SYS_STATVFS_H to the os.h file for Linux, as it has had this
+    support for a long time. Removed HAVE_SYS_VFS_H.
+
+18. Updated exipick to current release from John Jetmore.
+
+19. Allow an empty sender to be matched against a lookup in an address list.
+    Previously the only cases considered were a regular expression, or an
+    empty pattern.
+
+20. Exim went into a mad DNS lookup loop when doing a callout where the
+    host was specified on the transport, if the DNS lookup yielded more than
+    one IP address.
+
+21. The RFC2047 encoding function was originally intended for short strings
+    such as real names; it was not keeping to the 75-character limit for
+    encoded words that the RFC imposes. It now respects the limit, and
+    generates multiple encoded words if necessary. To be on the safe side, I
+    have increased the buffer size for the ${rfc2047: expansion operator from
+    1024 to 2048 bytes.
+
+22. Failure to deliver a bounce message always caused it to be frozen, even if
+    there was an errors_to setting on the router. The errors_to setting is now
+    respected.
+
+23. If an IPv6 address is given for -bh or -bhc, it is now converted to the
+    canonical form (fully expanded) before being placed in
+    $sender_host_address.
+
+24. Updated eximstats to version 1.33
+
+25. Include certificate and key file names in error message when GnuTLS fails
+    to set them up, because the GnuTLS error message doesn't include the name
+    of the failing file when there is a problem reading it.
+
+26. Expand error message when OpenSSL has problems setting up cert/key files.
+    As per change 25.
+
+27. Reset the locale to "C" after calling embedded Perl, in case it was changed
+    (this can affect the format of dates).
+
+28. exim_tidydb, when checking for the continued existence of a message for
+    which it has found a message-specific retry record, was not finding
+    messages that were in split spool directories. Consequently, it was
+    deleting retry records that should have stayed in existence.
+
+29. eximstats updated to version 1.35
+    1.34 - allow eximstats to parse syslog lines as well as mainlog lines
+    1.35 - bugfix such that pie charts by volume are generated correctly
+
+30. The SPA authentication driver was not abandoning authentication and moving
+    on to the next authenticator when an expansion was forced to fail,
+    contradicting the general specification for all authenticators. Instead it
+    was generating a temporary error. It now behaves as specified.
+
+31. The default ordering of permitted cipher suites for GnuTLS was pessimal
+    (the order specifies the preference for clients). The order is now AES256,
+    AES128, 3DES, ARCFOUR128.
+
+31. Small patch to Sieve code - explicitly set From: when generating an
+    autoreply.
+
+32. Exim crashed if a remote delivery caused a very long error message to be
+    recorded - for instance if somebody sent an entire SpamAssassin report back
+    as a large number of 550 error lines. This bug was coincidentally fixed by
+    increasing the size of one of Exim's internal buffers (big_buffer) that
+    happened as part of the Exiscan merge. However, to be on the safe side, I
+    have made the code more robust (and fixed the comments that describe what
+    is going on).
+
+33. Some experimental protocols are using DNS PTR records for new purposes. The
+    keys for these records are domain names, not reversed IP addresses. The
+    dnsdb PTR lookup now tests whether its key is an IP address. If not, it
+    leaves it alone. Component reversal etc. now happens only for IP addresses.
+    CAN-2005-0021
+
+34. The host_aton() function is supposed to be passed a string that is known
+    to be a valid IP address. However, in the case of IPv6 addresses, it was
+    not checking this. This is a hostage to fortune. Exim now panics and dies
+    if the condition is not met. A case was found where this could be provoked
+    from a dnsdb PTR lookup with an IPv6 address that had more than 8
+    components; fortuitously, this particular loophole had already been fixed
+    by change 4.50/55 or 4.44/33 above.
+
+    If there are any other similar loopholes, the new check in host_aton()
+    itself should stop them being exploited. The report I received stated that
+    data on the command line could provoke the exploit when Exim was running as
+    exim, but did not say which command line option was involved. All I could
+    find was the use of -be with a bad dnsdb PTR lookup, and in that case it is
+    running as the user.
+    CAN-2005-0021
+
+35. There was a buffer overflow vulnerability in the SPA authentication code
+    (which came originally from the Samba project). I have added a test to the
+    spa_base64_to_bits() function which I hope fixes it.
+    CAN-2005-0022
+
+36. The daemon start-up calls getloadavg() while still root for those OS that
+    need the first call to be done as root, but it missed one case: when
+    deliver_queue_load_max is set with deliver_drop_privilege. This is
+    necessary for the benefit of the queue runner, because there is no re-exec
+    when deliver_drop_privilege is set.
+
+37. Caching of lookup data for "hosts =" ACL conditions, when a named host list
+    was in use, was not putting the data itself into the right store pool;
+    consequently, it could be overwritten for a subsequent message in the same
+    SMTP connection. (Fix 4.40/11 dealt with the non-cache case, but overlooked
+    the caching.)
+
+38. Sometimes the final signoff response after QUIT could fail to get
+    transmitted in the non-TLS case. Testing !tls_active instead of tls_active
+    < 0 before doing a fflush(). This bug looks as though it goes back to the
+    introduction of TLS in release 3.20, but "sometimes" must have been rare
+    because the tests only now provoked it.
+
 
 Exim version 4.43
 -----------------