2004/09/12 == Released 0.8.19 == 2004/09/01 - Finished updating the code by adjusting to postfix-2.2-20040829 and started using it at my own site. 2004/08/01 - Started adjusting the patch to postfix-2.2-20040729. 2004/06/21 == Re-released 0.8.18 == 2004/06/21 - Postfix 2.1.3 has been released. Shortlived 2.1.2 did bring an incompatibel change (patch conflict) which has been resolved. - Fixed some typos in the tlsmgr.8 manual page (Chris Pepper ). 2004/04/27 == Re-released 0.8.18 == 2004/04/27 - Postfix 2.1.0 has been released. Some minor patch conflicts with respect to the actual code and build environment. - Due to the restructuring of the documentation the old sample-*.cf files are no longer available. Took documentation already adopted by Wietse for the 2.1-RC2-IPV6+TLS snapshot. 2004/02/09 == Re-released 0.8.18 == 2004/02/09 - Postfix 2.0.18-20040205 is available, patchkit applies without problems. 2004/02/02 == Release 0.8.18 == 2004/02/02 - Incorporated Luca Berra's information into the patchkit and ran tests with my own versions. 2004/02/01 - Reports about server side SMTP failure with Carsten's patch can be found on postfix-users. 'Luca Berra' informs, that he discoverd another failure of the GNU patch program with a misplaced patch hunk in smtpd.c 2004/01/30 - Edited in additional #ifdef USE_TLS conditionals. If the TLS patch is applied but not activated (USE_TLS is not defined), a warning is printed as soon as TLS shall be used. 2004/01/23 - Postfix 2.0.18-20040122 is now available. Several patch conflicts occur. Even more: one hunk of the patch (which is provided in unified diff) fails in smtp.c and causes a segmentation violation. Carsten Hoeger provides an adapted patch kit. 2004/01/02 == Released 0.8.17 == 2004/01/02 - Postfix-2.0.16-20031231 is released. No patch conflicts. - Changed autoresponder for TLS tests to "The Postfix Book" echo responder (provided by Patrick Koetter and Ralf Hildebrandt). 2003/12/30 - Postfix-2.0.16-20031226 is released. No patch conflicts. 2003/12/26 - Postfix-2.0.16-20031224 is released. Resolved patch conflicts. 2003/12/16 - Postfix-2.0.16-20031215 is released. Resolved patch conflicts. - src/global/pfixtls.c: changed occurance of "ssize_t" to "size_t" as some quite old operating systems do no have ssize_t (Reported by Klaus Jaehne for SunOS 4.1.4). - src/global/pfixtls.c: both the client and the server engine did print out messages even when tls_loglevel was set to 0 (reported by Florian Effenberger ): evaluate loglevel before printing any message. 2003/11/17 == Re-released 0.8.16 == 2003/11/17 - Postfix 2.0.16-20031113 is released. Some minor patch conflicts. 2003/10/27 == Re-released 0.8.16 == 2003/10/24 - Postfix 2.0.16-20031022 is released. Some minor patch conflicts. 2003/09/23 == Re-released 0.8.16 == 2003/09/23 - Postfix 2.0.16 and 2.0.16-20030921 are now available. Resolved some minor patch conflicts. 2003/09/10 == Released 0.8.16 == 2003/09/09 - Postfix 2.0.15 has been released including another workaround for select() on Solaris problems. It contains additional code to catch EAGAIN on read() in the timed_read() routine (and the respective precautions in timed_write() - Note: this fix is not yet part of Postfix 2.0.14-20030812. - Added corresponding code to pfixtls_timed_read()/_write(). - Changed SSL wrappermode behaviour: use smtpd_sasl_tls_security_options instead of smtpd_sasl_security_options as is to be expected because TLS is active. (Bug reported by Bob Snyder .) 2003/08/29 == Re-released 0.8.15 == 2003/08/29 - Adapted patchkit to Postfix 2.0.14. No patch conflicts. 2003/07/17 == Re-released 0.8.15a (-20030715 only) == 2003/07/16 - Experimental version Postfix 2.0.14-20030715 is released, including the SASL changes. Resolved some minor patch conflicts. 2003/07/11 == Released 0.8.15a (-20030706 only) == 2003/07/11 - Received error report about about TLS failing with the new smtpd_proxy feature including instructions on how to reproduce. (Did receive an earlier report on 2003/07/09, that however indicated other setup problems, so that the actual problem was not visible.) - Analysis: when introducing the new smtpd_proxy feature, different mechnisms where introduced to either write to the cleanup daemon (as before) or to the smtpd_proxy connection. Functions and streams are now expressed in out_fprintf() function pointers etc. being assigned accordingly. When updating to 0.8.15/2.0.13-20030706 this change was missed and the routine adding the TLS information to the Received: headers did use the older rec_fprintf() functions etc. This did work fine for the traditional connection to the cleanup service, but naturally failed for smtpd_proxy (with a segmentation violation). Solution: access out_stream via the according pointers. - The 2.0.13 stable version is not affected. 2003/07/08 == Released 0.8.15 == 2003/07/07 - Postfix 2.0.13 and 2.0.13-20030706 are released. Patchkit for 2.0.13 applies cleanly. Patchkit for 2.0.13-20030607 requires several adaptations (patch conflicts, no functional changes). - Slightly modified SASL interface code (smpt[d]_sasl_glue layer) to allow setting the security policy during session setup instead of process start. This allows to actually choose SASL mechanisms available depending on the availability of TLS encryption and authentication. New parameters: smtpd_sasl_tls_security_options, smtp_sasl_tls_security_options, smtp_sasl_tls_verified_security_options - Submitted change to SASL interface to Wietse, who accepted the change as part of the Snapshot line. 2003/06/19 == Released 0.8.14 == 2003/06/19 - Add support for SubjectAlternativeName "dNSName" entries in certificate checking (applies for client mode only). If the client connects to the server, it does check the list of dNSName entries against the expected hostname (therefore allowing the server to have multiple identities). As described in RFC2818 (HTTP over TLS), CommonName (CN) entries are only checked, if no dNSName entries are found at all. Initial patch proposed by Deti Fliegl , reworked to follow the RFC2818 rules and some cleanup. 2003/06/18 - Checked out similar settings, found another missing entry: var_smtp_scert_vd was missing src/smtp/smtp.c. - Renamed HAS_SSL to USE_TLS for compilation (have to use -DUSE_SSL in the future). Currently pfixtls.h will take care of setting USE_TLS, when HAS_SSL has been defined. 2003/06/17 - Received bug reports about Postfix/TLS failing (connection closing) after having finished the "STARTTLS"/"220 Ready to start TLS" dialogue. (Actually the first report came in via private mail on 2003/06/12, but the information was too diffuse to track down). Tracking down became possible after it became clear, that only Solaris systems are affected. Analysis: * As of 2003/06/09 postfix uses non-blocking socket I/O for the SMTP connection on Solaris platforms. This requires using "select()" style waiting before read() or write() access (which are not prepared EAGAIN or EWOULDBLOCK in the Postfix case and therefore indicate error). * As the var_smtpd_starttls_tmout variable is not correctly initialized (value is 0), the select() style function is not called, therefore read() fails with EAGAIN and the connection is closed due to a presumed error condition. * The initialization of the variable should be done in the time_table[] list during main(). The entry however was lost during the patch adaptation from 0.7.13e to 0.7.14-snap20020107 on 2002/01/07. Impact: * On Solaris systems, STARTTLS fails during handshake (server only). * On other systems, the TLS negotiation phase is not protected by the smtpd_starttls_tmout (default 300s) value and may hang until the watchdog kills smtpd, if the client does not continue the handshake. Restored var_smtpd_starttls_tmout variable initialization. 2003/06/12 == Re-released 0.8.13 == 2003/06/11 - Adapted to snapshot 2.0.12-20030611. No patch conflicts. 2003/06/11 - Adapted to snapshot 2.0.11-20030609. One minor patch conflict. 2003/05/23 == Re-released 0.8.13 == 2003/05/23 - First release against snapshot 2.0.10-20030523. 2003/04/26 == Re-released 0.8.13 == 2003/04/26 - Updated patchkit to apply to Postfix 2.0.9. - Updated patchkit-name to reflect the release of OpenSSL 0.9.7b. 2003/03/06 == Re-released 0.8.13 == 2003/03/06 - Postfix 2.0.6 has been released. No patch conflicts. 2003/03/02 == Re-released 0.8.13 == 2003/03/02 - Postfix 2.0.4 has been released. "patch" should work with some warnings about moved line numbers. - OpenSSL 0.9.7a has been released. No visible changes with respect to Postfix/TLS. 2003/01/26 == Re-released 0.8.13 == 2003/01/26 - Postfix 2.0.3 has been released. One minor patch-conflict. 2003/01/13 == Released 0.8.13 == 2003/01/13 - Postfix 2.0.1 has been released. Some minor patch conflicts resolved. - Added HOWTO documents contributed by Justin Davies to the contribution area. - Added RFC3207 (SMTP Service Extension for Secure SMTP over Transport Layer Security) to the documentation. RFC3207 is the successor of RFC2487. - Updated TODO list to reflect release ideas up to the release of Postfix/TLS 0.9.0. (Or will it finally be 1.0.0? :-) 2002/12/30 - OpenSSL 0.9.7 has been released. Postfix/TLS works best with the new 0.9.7 release. 2002/12/24 == Re-released 0.8.12 == 2002/12/24 - Postfix 2.0.0.1 has been released. Resolved one minor patch conflict. 2002/12/20 == Re-released 0.8.12 == 2002/12/20 - Postfix snapshot 1.1.12-20021214 has been released. Resolved minor patch conflicts. 2002/12/15 == Re-released 0.8.12 == 2002/12/15 - Postfix snapshot 1.1.12-20021214 has been released. Two minor patch conflicts. 2002/12/06 == Released 0.8.12 == 2002/12/06 - OpenSSL 0.9.6h has been released. Update documentation and filenames to reflect this new release. - Minor bug fix: when calling "sendmail -bs", smtpd is not run with superuser permissions, therefore the loading of the private key fails. STARTTLS is not used anyway, so the key is not needed anyway, but the failure to load creates a misleading warning. Do not initialize TLS engine at all when not started with superuser permissions. 2002/12/03 - Postfix snapshot 1.1.12-20021203 has been released. Resolved one patch conflict. 2002/11/01 == Re-released 0.8.11a == 2002/11/01 - Postfix snapshot 1.1.11-20021031 has been released. No patch conflicts. 2002/10/30 == Re-released 0.8.11a == 2002/10/30 - Postfix snapshot 1.1.11-20021029 has been released. No patch conflicts. 2002/09/30 == Re-released 0.8.11a == 2002/09/30 - Postfix snapshot 1.1.11-20020928 has been released. No patch conflices. 2002/09/24 - Postfix snapshot 1.1.11-20020923 has been released. Adapt patchkit. 2002/09/19 == Re-released 0.8.11a == 2002/09/18 - Postfix snapshot 1.1.11-20020917 has been released. Adapt patchkit. 2002/08/23 == Re-released 0.8.11a == 2002/08/23 - Postfix snapshot 1.1.11-20020822 has been released. Adapt patchkit. 2002/08/20 - Postfix snapshot 1.1.11-20020819 has been released with several enhancements and changes. Adapt patchkit (minor issues). 2002/08/12 - OpenSSL has experienced several (security critical) updates. 2002/07/26 == Re-released 0.8.11a == 2002/07/26 - On popular demand, a new diff for the snapshot version of Postfix is created: postfix-1.1.11-20020719. 2002/06/18 == Re-released 0.8.11a == 2002/06/18 - On popular demand, a new diff for the snapshot versions of Postfix is created: postfix-1.1.11-20020613. 2002/06/03 == Released 0.8.11a == 2002/06/03 - When compiling with SSL but without SASL, compilation fails due to the modification of state->sasl_mechanism_list that is not part of the "state" structure when SASL is not compiled in. This bug was introduced in version 0.8.11. Bug reported and patch supplied by Bernd Matthes . 2002/05/29 == Released 0.8.11 == 2002/05/29 - Postfix 1.1.11 is released. 2002/05/25 - Fix processing of options after STARTTLS handshaking: AUTH= was not handled, as the "=" was not recognized as for the extension list for the case without TLS. (The TLS case was a copy of an older version of the code not yet containing the "=" and the change in the main code slipped through without noting the difference, hence the option as not added to the TLS part. Found by "Christoph Vogel" . 2002/05/24 - Bug reported by "Christoph Vogel" : Client side AUTH does not work, if STARTTLS is used: if a server announces AUTH and STARTTLS, AUTH is being used if TLS is disabled. Once TLS is enabled, AUTH is still offered by the server, but the client does not use it any longer. Reason: when AUTH is offered, not only the SMTP_REATURE_AUTH flag is set in state->features, but also the available mechanisms are remembered in state->sasl_mechanism_list. As AUTH may be offered twice by some hosts (in the correct "AUTH mech" form and the older and deprecated "AUTH=mech" form), a check against processing the line twice is included in smtp_sasl_helo_auth(). This check now prevented the correct processing in the second evaluation of the ESMTP extensions offered after the STARTTLS activation. Solution: reset state->sasl_mechanism_list before processing the extension list just like state->features. 2002/05/15 == Released 0.8.10 == 2002/05/15 - Postfix 1.1.10 has been released. No changes. 2002/05/14 == Released 0.8.9 == 2002/05/14 - Postfix 1.1.9 has been released. Patchkit requires a small adjustment (supplied by Tuomo Soini ). 2002/05/10 == Released 0.8.8 == 2002/05/10 - OpenSSL 0.9.6d has been released. Release the unchanged patchkit with a new version number and under a new filename to indicate that it should be built against 0.9.6d (it has the session caching failure of 0.9.6c fixed). Update documentation accordingly. 2002/05/05 - Postfix 1.1.8 has been released, the patchkit applies cleanly. 2002/04/03 == Re-released 0.8.7 == 2002/04/03 - Postfix 1.1.7 has been released, the patchkit applies cleanly. Re-released the patchkit. 2002/03/29 == Released 0.8.7 == 2002/03/29 - Postfix/TLS did not honor the per-recipient-switching-off in SMTP client mode via tls_per_site (per-host-switching off was honored). Patch by Will Day . 2002/03/27 == Released 0.8.6 == 2002/03/27 - Postfix 1.1.6 has been released. Adapted patchkit to resolve minor patch conflict. (Template provided by Simon Matter ) 2002/03/13 == Released 0.8.5 == 2002/03/13 - Postfix 1.1.5 has been released. The patchkit would apply cleanly, but obviously the "lock_fd" change that applies to dict_dbm.c (Wietse) also has to be applied to dict_sdbm.c. Tuomo Soini kindly provided this change. 2002/02/25 == Released 0.8.4 == 2002/02/25 - Postfix 1.1.4 became visible. One patch conflict in a Makefile (Carsten Hoeger ). 2002/02/21 - Dates in this CHANGES document were showing 2001 even though 2002 already began :-). Fixed. (Marvin Solomon ) 2002/02/07 - Bug in the documentation (setup.html): the main.cf variables for the SMTP server process have to be named smtpd_* instead of smtp_*. Found by Andreas Piesk . 2002/02/03 == Released 0.8.3 == 2002/02/03 - Patch from Andreas Piesk : remove some compiler warnings by using explicit type casts in hexdump print statements. - Re-released otherwise unchanged patchkit against Postfix-1.1.3. 2002/01/30 == Released 0.8.2 == 2002/01/30 - Re-released unchanged patchkit against Postfix-1.1.2. 2002/01/24 == Released 0.8.1 == 2002/01/24 - Postfix-1.1.1 has been released. The patchkit needed some small adjustment. - Both Tuomo Soini and Carsten Hoeger helped out with this small adjustment. As a side effect of Carsten's complete pfixtls.diff, which I compared after applying Tuomo's adjustment, I found that pfixtls.c contained several wrong "'" characters: on the german keyboard there is an accent looking like the apostroph but producing a different binary code. Obviously on Carsten's machine the code was changed which became obvious during the comparison. (Conclusion: I wrote the comments affected on my SuSE-Linux PC at home with german keyboard. In my university-office I do have HP-UX workstations with US keyboards.) 2002/01/22 == Released 0.8.0 == 2002/01/22 - Received a comment from Wietse on the mailing list, that it is better to resolve the "standalone" issue by using the already available SMTPD_STAND_ALONE() macro in smtpd. Undid 0.7.16 change and made new change in smtpd.c. - Updated links in the References section of the documentation. 2002/01/21 == Released 0.7.16 == 2002/01/21 - When calling "sendmail -bs" and STARTTLS is enabled, smtpd tries to read the private key and fails due to insufficient permissions (smtpd is run with the privileges of the user). This case is caught since version 0.6.18 of the Postfix/TLS patchkit: STARTTLS is still being offered but a "465 temporary failure" message is issued. Some mailers (read this: PINE) will then refuse to continue. (And an irritating error message indicating the failure to read the key will be logged.) Experienced by "Lucky Green" . - Solution: Disable STARTTLS when running "sendmail -bs" by adding "-o smtpd_use_tls=no -o smtpd_enforce_tls=no" to smtpd's arguments upon startup. Using STARTTLS does not make sense in simulated SMTP mode. 2002/01/18 == Released 0.7.15 == 2002/01/18 - Postfix 1.1.0 has been released. The patchkit for the former snapshot version applied cleanly and now becomes the patchkit for the stable version. 2002/01/16 == Released 0.7.14a == 2002/01/16 - Snapshot-20020115 is released. Adapted patchkit. - Add Postfix/TLS entries into the new conf/postfix-files (Tuomo Soini , Carsten Hoeger ). 2002/01/14 - OpenSSL: a user reported that session caching stopped working for him with OpenSSL 0.9.6c. I found that this is also true for my own Postfix/TLS installation. Solution: server side session caching is broken in OpenSSL 0.9.6c when using non-blocking semantics (Postfix/TLS is affected as it uses BIO-pairs); sessions are simply not added to the cache. This bug is not security relevant. A fix has been applied to the OpenSSL source tree for the next release. 2002/01/08 == Released 0.7.14 == 2002/01/07 - New snapshots released as release candidates. Adapted the patchkit to snapshot-20020107. Moved our production servers from 20010228-pl08 to snapshot-20020107 with the adapted patchkit. - Fix documentation: tlsmgr can be run chrooted since a long time. 2001/12/21 - OpenSSL 0.9.6c is released. Postfix/TLS is fully compatible. 2001/12/19 == Released 0.7.13e == 2001/12/19 - Adapted patchkit to snapshot-20011217. 2001/12/12 == Released 0.7.13d == 2001/12/12 - Adapted patchkit to snapshot-20011210. Adaption provided by Tuomo Soini . 2001/11/28 == Released 0.7.13c == 2001/11/28 - Adapted patchkit to snapshot-20011127. 2001/11/26 == Released 0.7.13b == 2001/11/26 - Adapted patchkit to snapshot-20011125. 2001/11/22 == Released 0.7.13a == 2001/11/22 - Adapted patchkit to snapshot-20011121. 2001/11/15 == Released 0.7.13 == 2001/11/15 - Adapted patchkit to postfix-20010228-pl08 and snapshot-20011115. 2001/11/06 == Re-released 0.7.12 == 2001/11/06 - Snapshot-20011105 released. No patch conflicts, but in order to have the pfixtls-* filename and home page entry reflect the new version, I'll re-release 0.7.12. 2001/11/05 == Released 0.7.12 == 2001/11/05 - Release of Postfix-20010228-pl06 and snapshot-20011104. The snapshot version had some minor patch conflicts to be resolved. 2001/10/14 == Released 0.7.11 == 2001/10/14 - Bug fix (client mode): when the peername is checked against the CommonName in the certificate, the comparison does not correclty ignore the case (the peername as returned by DNS query or set in the transport map is not transformed to lower case). This bug was introduced in 0.7.5. 2001/10/09 == Released 0.7.10 == 2001/10/09 - Snapshot-20011008 is released. Some minor adaptions are required to sort out patch conflicts. 2001/09/28 - Received patch from Uwe Ohse : There is a bug in sdbm's handling of the .dir file, that also applies to Postfix/TLS. The problem only appears for large databases. - The example entries in conf/master.cf for the submission and smtps services use "chroot=y" flags, while the Postfix default is "chroot=n". This could lead to hardly explainable problems when users did not note this fact during setup. Fixed example entries to also use "chroot=n" default. 2001/09/18 - Wietse releases Postfix-20010228-pl05. The patch applies cleanly with "patch -p1 ...", so it is not necessary to release a new patchkit. 2001/09/04 == Released 0.7.9 == 2001/09/04 - Due to unititialized variable in smtpd_state.c, AUTH may not be offered without TLS even though smtpd_tls_auth_only was not enabled. (Patch from Nick Simicich .) 2001/08/29 - In the snapshot-20010808 version of 0.7.9, the "tlsmgr" line in the sample conf/master.cf is missing (reported by Will Day ). Fixed. 2001/08/27 == Released 0.7.8 == 2001/08/27 - Received bugreport about issuer_CN imprints consisting of long strings of nonsense. This only appeard with certificates issued from a certain CA (RSA Data Security Inc., Secure Server Certification Authority). (Will Day ) - The problem: the issuer data of this certificate is: Issuer C=US O=RSA Data Security, Inc. OU=Secure Server Certification Authority It does not contain a CN (CommonName) field. OpenSSL's X509_NAME_get_text_by_NID() function does not catch this condition (no error flag set), but it also does not set the name in the memory location specified. - Solution: 1. Preset the memory for the string to '\0', so that a string of length 0 is obtained on the failure described above. 2. When no CN data is available, use the O (Organization) field instead. The data are used for logging only (it is the issuer, not the subject name), so this change does not affect functionality. 2001/08/22 == Released 0.7.7 == 2001/08/22 - Found one more bug: erronously called SSL_get_ex_new_index() instead of SSL_SESSION_get_ex_new_index() (note the _SESSION missing). This could be responsible for the failure at the locations found during debugging. Works fine on HP-UX (did also before), must cross check at home... 2001/08/21 - Received report, that smtp (client) fails with signal 11 (platform: linux redhat). Cannot reproduce any problem on HP-UX (did run 1 week in production before release). But malloc() and stack strategies are different between platforms. - Can reproduce the failure on my Linux PC at home :-(. - Found one bug in new_session_cb(): on successfull external caching, success is reported by a return value of 1. This however must be another bug, as it has nothing to do with the locations of the failure, when analyzing the core dumps/running under debugger. Still getting SIGSEGV... 2001/08/20 == Released 0.7.6 == 2001/08/20 - Following "popular demand" implemented new feature and configuration option "smtpd_tls_auth_only": Only allow authentication using the AUTH protocol, when the TLS encryption layer is active. Default is "no" in order to keep compatiblity to postfix without TLS patch. This option does not distinguish between different AUTH mechanisms. 2001/08/16 == Released 0.7.5 == 2001/08/15 - The new session cache handling is working now at my site for quite some time. - Client side: modified peername matching code, such that wildcard certificates can be used. Matching is done as in HTTP/TLS: only the leftmost part of the hostname may be replaced by a '*'. 2001/08/09 - Further debugged the CRYPTO_set_ex_data() functionality. - Unified "external cache write" and "external cache remove" callbacks for client and server side. The "external cache read" functions are not that easy to combine, as the lookup keys are quite different and do not match the fixed interface to the callback function. - Change shutdown behaviour according to SSL_shutdown(). When SSL_shutdown() returns, the shutdown handshake may not be complete, if we were the first party to send the shutdown alert. We must call SSL_shutdown() again, to wait for the peer's alert. 2001/08/08 - Postfix snapshot 20010808 is being released. 2001/08/08 - Rewrite server side to remove externally cached sessions via callback. - Rewrite client side to remove externally cached sessions via callback. This turns out to be more difficult as expected, as the client side session cache is sorted by hostnames, but the callbacks are called with the SSL_SESSION objects. The information must be stored into the SSL_SESSION objects by using the CRYPTO_set_ex_data() functionality, the documentation of which, ahem, ... - Reloading sessions stays separate, as the functionality is different. 2001/08/07 - Started reworking the session cache code. * On the server side the retrieval from the external cache and the writing to the cache are handled by callback functions. The removal is handled directly. * On the client side, all session cache operations are performed explicitly. * The explicit handling is on the client side is bad, as it requires a quite complicated logic to detect session reuse and the appropriate handling. * The explicit handling of session removal on both sides is bad, as the OpenSSL library will remove sessions (on session failure) according to the TLS specifications automatically, so we want to take advantage of this feature and have the externally cached sessions removed as required via callback. - First step: on the client side, also use the new_session_cb(), so that new sessions are automatically saved to the external cache on creation. 2001/08/01 - Postfix-20010228-pl04 is being released. 2001/07/11 == Released 0.7.4 == 2001/07/10 - Postfix snapshot 20010709 was released. Resolved some minor patch conflicts. 2001/07/10 - OpenSSL 0.9.6b has been released including a security fix for the libraries internal pseudo random number generator. * Note: to exploit the weakness, an attacker must be able to retrieve single random bytes. As in Postfix/TLS random bytes are only used indirectly during the SSL handshake, an attacker could never access the PRNG in the way required to exploit the weakness. * Postfix/TLS is therefore not vulnerable (as are most (all?) applications utilizing the SSL layer). * The OpenSSL team however recommends to upgrade or install the bugfix included in the announcement in any case. * Details can be found at http://www.openssl.org/ 2001/05/31 == Released 0.7.3a == 2001/05/30 - Report from : TLS logging does not work. Reason: parameters are not evaluated in mail_params.c, as the corresponding lines for other_int_defaults[] were missing from the patch. This only affected the 0.7.3-snapshot version, the version for "stable" is correct. I will release 0.7.3a with this fix only for the snapshot version to keep version numbering consistent with the "stable" version. 2001/05/28 == Released 0.7.3 == 2001/05/28 - Upgraded to snapshot-20010425: resolved some minor patch conflicts. No functional changes. 2001/05/16 - Received french documentation (doc_french/) contributed by Etienne Roulland . 2001/05/03 == Released 0.7.2 == 2001/05/03 - Postfix-Snapshot 20010502 is released. Bernhard Rosenkraenzer supplies an adapted patch for Postfix/TLS, as the normal patch has several rejections because of code changes; functionality has not changed. 2001/05/01 - Patchlevel 02 of Postfix 20010228 is being released. The Postfix/TLS patchkit applies cleanly when using the "-p1" switch to patch. 2001/04/09 == Released 0.7.1 == 2001/04/06 - OpenSSL 0.9.6a is released. It contains several bugfixes and will become the recommended version to be used with Postfix/TLS. I will run some more test and then re-release Postfix/TLS (without additional changes to the source) as 0.7.1 to make people aware of the new versions of Postfix and OpenSSL. 2001/04/05 - Hint from Bodo Moeller : the "Known Bugs" section in doc/test.html actually contains bugs of clients and/or interoperatbility problems. Better name it "Known interoperability problems" and rename the entries "Postfix/TLS server" and "Postfix/TLS client" to improve clarity. 2001/03/29 - Patchlevel 01 of Postfix 20010228 is being released. The Postfix/TLS patchkit applies cleanly when using the "-p1" switch to patch. OpenSSL 0.9.6a will be out within the next handful of days, so I will delay the release of a new patchlevel until then. 2001/03/01 == Released 0.7.0 == - IMPORTANT: If you are upgrading from a much older version, you will find that some configuration options have changed over time (fingerprints are now handled with ':'. check_relay_ccerts is now permit_tls_clientcerts. Session caching has been reworked.) It is recommended to re-read the sample-tls.cf file or the html version in the documentation. 2001/03/01 - Wietse has announced the _release_ version (non-beta) or postfix: 20010228! - Applied the Patchkit to the _release_ version (not the snapshot version). Resolved one minor patch conflict. - So, it's time to call this Postfix/TLS 0.7.0. 2001/02/26 == Released 0.6.38 == 2001/02/26 - Snapshot-20010225 has been released. Resolved one minor patch conflict. 2001/02/23 == Released 0.6.37 == 2001/02/23 - Snapshot-20010222 has been announced as RELEASE CANDIDAT. Resolved one minor patch conflict. - Removed "check_relay_ccerts" restriction which has been replaced by "permit_tls_clientcerts" in 0.6.24. (Was left in until now for transition.) - Do not try to save session data > 8kB, since this cannot be handled by SDBM. (This is more or less academical, since I have never met a session even half that large.) 2001/02/19 == Released 0.6.36 == 2001/02/05 - Snapshot-20010204 has been released. Resolved one minor patch conflict. 2001/02/03 == Released 0.6.35 == 2001/02/03 - Snapshot-20010202 has been released. Resolved one minor patch conflict. 2001/01/29 == Released 0.6.34 == 2001/01/29 - Snapshot-20010128 has been released. Resolved some minor patch conflicts. 2001/01/11 == Released 0.6.33 == 2001/01/10 - Discussion in Thread "When to get peer certificate?" continues and it comes out, that cross references between datastructures are well maintained inside OpenSSL. A fact not well known due to lack of documentation (seems I am facing some more work on the OpenSSL manpages :-). - Moved around data needed for the certificate verification: a lot of "static" entries globally needed inside pfixtls.c could now be moved into the connection specific TLScontext. 2001/01/07 == Released 0.6.32 == 2001/01/07 - Since now the checks at handshake stage (in pfixtls.c) are more strict, some of the checks in smptd.c and smtp_proto.c could be removed. At a later point I can probably move even more checks into pfixtls.c... 2001/01/05 - Had a discussion with Ari Pirinen on openssl-users (Thread: When to get peer certificate?) about the earliest possible place to check the CommonName of the peer against the expected name. (This is what smtp does when enforcing the peername of the server it is connecting to.) The final result was, that the check can already been done inside the verifiy_callback() routine even before the handshake is completed. The positive side effect is, that since the session is never completly established, it is also not cached on either client or server. - Since this is a good idea, I have extended the verify_callback in src/global/pfixtls.c to check the CommonName of the peer (if applicable) and have the handshake shut down immediatly on failure. I have also changed the behaviour so that whenever a positive certificate verification is required, the handshake is shut down immediatly. (The versions up to now did delay these checks until the session was established and then shut down the connection. I had established this practice while working on BIO-pairs and running into a bug in OpenSSL 0.9.5 (fixed now) and with the verify depth.) 2000/12/23 == Released 0.6.31 == 2000/12/23 - Bug: When only enabling smtpd_tls_wrappermode and not additionally setting smtpd_use_tls or smtpd_enforce_tls, the TLS engine was not fired up on startup of smtpd Fixed: also start TLS engine when only smtpd_tls_wrappermode is enabled. (Experienced by "Fiamingo, Frank" ) 2000/12/18 == Released 0.6.30 == 2000/12/18 - New snapshot 20001217 has been released. Due to the change of "timeout" parameters now being its own class and table, the old patchkit does not apply cleanly! - Checked out Postfix/TLS parameters being timeout values and put them into the new style time parameter table. This allows to specify time values like 3600s or 1h. Updated sample configuration to reflect this new style. - "Fiamingo, Frank" pointed out to me, that there are three parameters in src/global/mail_params.h (namely DEF_TLS_RAND_EXCH_NAME, DEF_SMTPD_TLS_CERT_FILE, DEF_SMTPD_TLS_CA_FILE) that are hardcoded as "/etc/postfix/something". This does not match the usual style of postfix, where no paths are hardcoded this way. I have removed the defaults for CERT_FILE and CA_FILE. The RAND_EXCH is needed for good PRNG seeding on systems without /dev/urandom, I however don't know yet, how to rearrange this requirement. I could use the Postfix internal mechanisms to enforce a parameter, but this would annoy people having compiled in TLS but not activated. 2000/12/13 == Released 0.6.29 == 2000/12/13 - Snapshot-20001212 has been released. - Undid bugfixes for 20001210 which now are included in the new snapshot. 2000/12/12 == Released 0.6.28 == 2000/12/12 - Added bugfix provided by Wietse on postfix-users@postfix.org for "postconf -m" behaviour. 2000/12/11 - New snapshot-20001210 released. Some patch conflicts occur. Additionally * adjusted calls to myflock() to changed interface, * fixed bug in smtpd_sasl_glue(), where a change to the name_mask() call was not applied in the original snapshot. 2000/12/05 == Released 0.6.27 == 2000/12/04 - Print informational message "SSL session removed" only when var_smtp[d]_loglevel >= 2. (Proposed by Craig Sanders .) - Extend logging of "setting up TLS connection from/to" and corresponding success/failure messages so that they include the hostname/ip address. This way it is much easier to automatically analyze errors by simply grepping for e.g. "SSL_accept error" and immediately get the peer causing the problem without further logfile processing. (Proposed by Craig Sanders .) - When experiencing a TLS failure due to TLS-enforced failure in client mode (no certificate or hostname/certificate mismatch etc), immediately shut down the TLS mode with "failure" indication, so that the SSL session is removed immediately. This way a new session is always enforced in the case the peer has fixed the problem; no need to wait for the timeout. 2000/11/29 == Released 0.6.26 == 2000/11/29 - Found security relevant bug in the OpenSSL library: the verify_result stating whether or not the certificate verification succeeded is not stored in the session data to be cached and reused. - This bug was found during the development of Postfix/TLS around one year ago, the bug in the library was however only fixed for the server side. At that time I also tested the server side behaviour but ommitted to check the client side, too. - Versions before Postfix/TLS 0.4.4 experienced this problem for both server and client side. Before 0.6.0 a workaround was active for both sides, which has been removed at 0.6.0 in the believe that the bug was gone (I only tested the server side, which was fixed). - Fixed that bug in OpenSSL also for the client side (I can do this myelf now that I have been invited to join the OpenSSL developers team :-). The fix is availabe as of today and will be part of the 0.9.7 release of OpenSSL (or 0.9.6a, if this release will be published). - Included a workaround inside Postfix/TLS for OpenSSL library versions before 0.9.6a or 0.9.7, respectively. ********************** Begin Description - By not caching the verify_result for the client side, the following behaviour could appear: * The problem can only appear when smtp_tls_session_cache_database is activated. * smtp_use_tls = yes X On the first connection, the certificate fails verification, failure is logged: smtp[*]: Unverified: subject_CN=serv01.aet.tu-cottbus.de, issuer_CN=BTU-CA For any following connections until the session times out (default 1 hour), the peer certificate seems to pass verification: smtp[*]: Verified: subject_CN=serv01.aet.tu-cottbus.de, issuer_CN=BTU-CA X Security Impact: Unverified certificates are logged as if verification had succeeded. * smtp_enforce_tls = yes X After the verification failure, the session is never correctly established and hence not reused. X Security impact: None, as the session is never reused. * smtp_enforce_tls = yes after smtp_tls_enforce_tls = yes for a server. X If the session has been recorded with use_tls and then for this server enforce_tls is set, the wrong verify_result could be used within the session cache timeout (default = 1 hour). X Security impact: If TLS shall be enforced for a recipient, there is a window of approx. one hour from setting the "enforce_tls" switch until a verification failure is noted. For this to happen, a TLS session to that server must have been used with use_tls set and the not-verifiable certificate must have been recorded in that session. - Evaluation: Even though this _is_ a security problem, I consider risk to be *low*, given the conditions under which the problem might occur. ********************** End Description 2000/11/27 == Released 0.6.25 == 2000/11/26 - Added "permit_tls_all_clientcerts" for smtpd_recipient_restrictions. When this option is enabled, any valid client certificate allows relaying. This can be practical, if e.g. a company has a special CA to create these certificates and only this CA is "trusted". It however does not allow finer control, so if e.g. an employee leaves, he could still relay. Postfix/TLS does not (yet) allow CRL (certificate revocation lists). (Added on popular demand.) - Make the client behaviour more configurabe: when enforcing TLS connections, the peer's name is checked against the CommonName in its certificate. New configuration variable "smtp_tls_enforce_peername" (default=yes) can now be used to accept peername!=CommonName. The server's certificate must still pass the verifcation process against a trusted CA! In tls_per_site, the according key is MUST_NOPEERMATCH. (Added on demand.) 2000/11/24 - If the server requires a client certificate and no certificate is presented or the certificate fails verification, the connection is shut down but no information is logged. -> add according msg_info() in smtpd/smtpd.c:startls_cmd(). - If TLS is not enforced, it does not make sense for a server to require a client certificate. If no STARTTLS is issued, the SMTP would continue anyway, so why shut down when TLS is activated without verifyable client certificate? -> ignore smtpd_tls_req_ccert=yes, if TLS is not enforced and only treat like smtpd_tls_ask_ccert = yes with an according information logged. 2000/11/22 == Released 0.6.24 == 2000/11/22 - Installed on my own servers and changed configuration to use the new "permit_tls_clientcerts" option name. Patchkit will be released after some hours of successfull operation. 2000/11/21 - New snapshot-20001121 is being released. The patch applies without any conflict when applied with "patch -p1", so no need to rush out an updated patchkit. - Rename the smtpd_recipient_restrictions option from "check_relay_ccerts" to "permit_tls_clientcerts" to better match the naming scheme. Leave in the old option for now to not break existing configurations. The final incompatible removing is scheduled of release 0.7.0 of the patchkit which will be matching the next "stable" release of postfix. - There is no manual page for tlsmgr.8 (pointed out by Terje Elde ). Fix the comments at the beginning of tlsmgr.c and create tlsmgr.8. - In the session cache code an additional 20 bytes were allocated when converting SSL_SESSION data to binary using i2d_SSL_SESSION(). In adding these 20 bytes to the size listed by i2d_SSL_SESSION() I followed the example in the OpenSSL source (PEM_ASN1_write()). These 20 bytes are only added since when writing the PEM, a 20 byte checksum is added, so we don't need it in our case -> removed. (Researched after Carlos Vicente asked what these 20 bytes are good for :-) 2000/10/30 == Re-Released 0.6.23 == 2000/10/30 - Postfix snapshot-20001030 with an important bug fix is made available. The patchkit applies without any problem (patch -p1). Hence, I re-release the 0.66.23 release for the new snapshot. 2000/10/30 == Released 0.6.23 == 2000/10/30 - New Postfix snapshot 20001029 available with some important bug fix. Adjusted patchkit (only minor conflicts). 2000/10/27 - The CN_sanitize function (src/smtpd/smtpd.c) that shall make sure that no illegal sign is included into the Received: header does not work on systems were "char" is unsigned by default. (Linux on s390, found by Carsten Hoeger ) -> Worked out a more precise (even though not looking elegant) solution that checks out all acceptable characters. - Sent new smptd.c to Carsten Hoeger for testing, will wait with new Postfix/TLS release. 2000/10/06 == Released 0.6.22 == 2000/10/06 - snapshot-20001005 has been released, featuring fast ETRN. Only some minor patch conflicts needed to be resolved. 2000/09/28 == Released 0.6.21 == 2000/09/28 - snapshot-20000924 seems to be somewhat longer lasting. I have been asked for a new Postfix/TLS release against snapshot-20000924, hence I will create one. - Running OpenSSL 0.9.6 for a week now to my full satisfaction. I will bump bump up the Postfix/TLS version counting to include "0.9.6", even though it will still run fine with 0.9.5a. 2000/09/25/ - snapshot-20000924 is available; only small adjustments. - Wietse seems to release new snaphots on a daily basis, it doesn't make sense to follow with a new Postfix/TLS release every day. 2000/09/23 == Released 0.6.20 == 2000/09/23 - Recompile OpenSSL-0.9.6-beta3 with the change and reinstall old pfixtls.c: works again. Hence, all versions of Postfix/TLS working against 0.9.5a will also work again 0.9.6-final, which shall be released on 2000/09/24! - Wietse releases snapshot-20000923, patchkit adapted. - Went through the "install.html" document to add a remark about OpenSSL-0.9.6. This document is of historic quality but did not fit actual versions of Postfix/TLS, we are far beyond OpenSSL 0.9.2: Updated. 2000/09/22 - Wietse releases snapshot-20000922. The source directory hierarchie has changed, so the patch needs to be adjusted at several places. - Run tests against OpenSSL 0.9.6-beta3: problems occur! * Certificates are no longer verified, since an informationa flag about the CA certificate search process is written into the error storage and thus misinterpreted as verification failure. * Changed Postfix/TLS source to maintain its own error storage based on the verify_callback, send out according warning to Postfix/TLS mailing list. * Unfortunately, this will break all older versions of Postfix/TLS. Sent out analysis to OpenSSL-bugs@openssl.org. * Additional change is made to OpenSSL: the new behaviour is only activated when a special flag is set, so compatibility is restored! 2000/09/21 - Wietse releases snapshot-20000921. Some minor patch conflicts resolved. 2000/09/14 == Released 0.6.19 == 2000/09/14 - Received a bug report: Postfix/TLS will accept a mail even though smtpd_req_ccert=yes (require use of client certificate) and no client certificate is presented. Reason: when no client certificate is presented SSL_get_verify_result() will return X509_V_OK, since this is the default value. Solution: only set "peer_verified" internal information, if the verify_result is X509_V_OK _and_ a peer certificate is available. Remark: This default value does not make too much sense. I will file a bug report/patch before the next release of OpenSSL... 2000/09/03 == Released 0.6.18 == 2000/09/03 - When calling "sendmail -bs", smtpd is started without root privileges, hence it cannot open the private key file and the session cache database. Since the database routines do not offer a graceful return (only fatal and abort), this leads to a failure when TLS and session caching is activated. This affects PINE users (noted by Craig Sanders ). Solution: Try to read the private key first; if that fails, we can gracefully recover and won't touch the session cache database at all. - When STARTTLS is configured for smtpd but does not work (e.g. because of unaccessible keys), smtpd answers with "465 TLS not available due to temporary reasons". After that the connection was closed, this is however not necessary, as the client may decide to continue without TLS activated. - Craig Sanders contributes a script to automatically generate the keys and certificates for Postfix/TLS usage. Added "make-postfix-cert.sh" to the contributed/ directory. 2000/09/02 == Released 0.6.17 == 2000/09/02 - Craig Sanders reports that he has connection problems with a site; the message in the log is: SSL_connect error 0 8847:error:140943F2:SSL routines:SSL3_READ_BYTES:sslv3 alert unexpected message:s3_pkt.c:956:SSL alert number 10: * This is the error caused by the faulty TLS implementation with CommunigatePro. The bug is fixed in later versions of CommunigatePro, The site shall be contacted, they should update. - More important, he reports a segmentation fault immediately after this problem. - Bug: when not using session caching and an error occurs during the TLS handshake, pfixtls_start_clienttls() tried to remove the erronous session from a non-existant session cache. Fix: check the existence of the session cache before trying to access it. Comment: at all other places in the code this condition was already caught. - Remark: actually session caching was configured, but the configuration variable was mistyped because... it was wrong in conf/sample-tls.cf and doc/conf.html. The correct values are "smtp[d]_tls_session_cache_database" instead of "smtp[d]_tls_use_session_cache_database". Unfortunately this is not flagged by Postfix... 2000/08/25 == Released 0.6.16 == 2000/08/25 - Make sure, that the smtp[d] processes will try to access the "daemon" entropy sources, but will only print an info when not available. Using the PRNG-exchange file, they can happily run without. - Moved HAS_SSL checks, such that the package compiles also when configured without -DHAS_SSL. 2000/08/24 - Changed the handling of the PRNG-exchange file. Until now it was written by tlsmgr and read by the smtp[d] daemons. This had the disadvantage, that until tlsmgr rewrote new bytes to the file, all starting daemons read the same seed (to which some more bits, but not too much were added). - Now the file is handled in read->stir into pool->write back mode, so that every daemon will add its own entropy bits. - The smtp[d] processes will do so when starting, when opening a TLS connection and when closing. - The tlsmgr will also read back the file and add it to its pool, so that no entropy is lost. - This change significantly increases the "self seeding" capability of the TLS service. 2000/08/09 - Cleaned up the new PRNG-seeding. - When tlsmgr looses connection to an EGD-source (because it was restarted), tlsmgr performes an exit(0), so that a newly started tlsmgr can reconnect. [chroot/dropped privileges]. 2000/08/04 - Introduced new entropy sources for single daemons: * tls_daemon_random_source Using this source (same style as for tlsmgr), each starting daemon can obtain additional entropy (32 bytes by default). The PRNG-exchange file is still read. - I am not sure about the policy for this feature. If such a source is given, should a failure be considered fatal? 2000/07/23 - Started reworking the PRNG seeding: * tlsmgr now recognizes tls_random_source as dev:/dev/urandom /* Direct read from device file */ egd:/path/to/socket /* Connection via EGD-socket */ /path/of/plain-file * If a dev: or egd: is given, tlsmgr will connect and keep the connection open, so that it now can run in chroot-mode with dropped privileges. - Since EGD can be drained, but the connection is permanently open, only suck a small number of bytes (default 32) at a time, but do it more often. 2000/08/09 == Released 0.6.15 == 2000/08/09 - Traced through OpenSSL to learn more about the verify_callback-feature. The callback is called several times. When it returns "1", the handshake will continue, when it calls "0", the handshake will immediately fail (and Postfix/TLS will also close the TCP connection). - Following the sample in the OpenSSL-apps, the verification chain depth was the only property triggering this effect, so this stood hidden until now. Obviously, users having longer chains did set the verifcation depth accordingly or they gave up, since this was never reported... - Changed the behaviour of verify_callback() to never return "0", such that we can deal with the verification result later in a more consistent manner. If we only enable and not enforce, we simply want to ignore problems with the certificate. - verify_callback() did not print out all information, since the wrong state variables (pfixtls_*active instead of pfixtls_*engine) were checked. The *active state variables are only set later. As the verify process now became rather narrative, the normal logging is only done in loglevel 2! - Arrrghhh. The conf/sample-tls.cf _and_ the html-docu (which is actually copied from conf/sample-tls.cf) has wrong names for the verification- depth parameters. *_vd instead of *_verifydepth and ccert<->scert. [Wondering, why this never popped up before...] - Changed the default-verifydepth to "5" which should suffice for most cases. Maybe the limit could also be completely removed, but we should at least receive a warning hint when something goes wild. Since OpenSSL>=0.9.5 is required for Postfix/TLS anyway, certificate chain verification can now be used, so the caution applied before is no longer necessary. 2000/08/08 - Tracked down the double-free() call in smtp with Efence. SSL_free() does call SSL_SESSION_free() on the negotiated session. Hence, I must not call SSL_SESSION_free() on the session in question, it will be removed anyway. - Also tracked down the certificate chain feature. Reason is the verify_callback() in global/pfixtls.c. It flags a chain depth that is too long as fatal, hence the connection is immediately closed. 2000/08/04 - Received information from Alain Thivillon : FreeBSD-CURRENT offers malloc() with additional checks enabled. After successfully delivering, smtp dumps core with free() called twice in TLS mode. - I noted, that there is a communication problem with his site an my new certificate issued by the universities computer center (which has a chain depth of 2). Step back to the old self certificate for the time being. 2000/07/27 == Released 0.6.14 == 2000/07/27 - Introduced new configuration parameter "smtpd_tls_wrappermode" that enables the (deprecated) old style SSL-wrapping around SMTP. It could be run on a different port (once smtps=465) was recommended for this services. This method is used by old versions of Outlook (Express), the Mac versions and even actual versions, when not run on port 25. [Actually it was only a handful of lines, so it doesn't hurt too much, even though it does not follow any RFC.] - I recommend using this option only from master.cf. Example lines added to conf/master.cf and description added to Postfix/TLS-doc/conf.html. - When having SASL enabled and TLS-enforce mode in "smtpd", only offer AUTH, when TLS has been activated. Otherwise the client might simply send the unencrypted credentials before it receives 530 Must issue a STARTTLS command first and an eavesdropper already has what he was looking for. 2000/07/19 == Released 0.6.13 == 2000/07/19 - Changed the library-initializaton call to new naming scheme (SSLeay_add_ssl_algorithms() to OpenSSL_add_ssl_algorithms() :-). - Updated documentation to reflect the use of chain certificates with CAfile and smtp[d]_tls_cert_file (see 2000/07/06). - Documentation: the interoperability problem with CommunigatePro has been solved: CommunigatePro violated the TLS-RFC and has been fixed. - Typo: It is "to stir" not "to stirl" :-) 2000/07/06 - Received certificate for our site from our computer center. It's a chain certificate. Now load the cert with SSL_CTX_use_certificate_chain_file(), in order to better load the chain CA certificates. 2000/07/04 - Reported Wietse about a possible problem in the SASL code, a relay check may also be performed if sasl was not enabled and might lead to unwanted relay. As the fix is in my own codebase, I will leave it Postfix/TLS until a new snapshot (or final release) is available. 2000/06/02 == Released 0.6.12 == 2000/06/02 - Adapted to Snapshot-20000531 (minor patch conflict). - Cleaned up some old header file dependencies in global/pfixtls.c and global/Makefile.in that are no longer needed due to the interface changes (timed_read()/write()) in 0.6.7. 2000/05/29 == Released 0.6.11 == 2000/05/29 - Following Bodo Moeller's analysis, the error is due to a mismatch between the CA certificate accessible in the smtp[d]_tls_CAfile and the one used in the actual certificate (smtp[d]_tls_cert_file). Daniel Miller fixed his setup and the problem is gone. - Introduced a workaround into Postfix/TLS: if the padding error is found, it is removed from the error-queue by Postfix/TLS, in order to protect more sites from experiencing this problem. - Added a warning to conf/sample-tls.cf - Updated to the latest snapshot-20000528. 2000/05/27 - After some fiddling around working through the binary certificate data to see where it is modified at 0.6.10, I actually note, that both 0.6.9 and 0.6.10 choke on the data. Now going back up through the functions very fast reveals the problem: * The certificate supplied triggers the "RSA-padding" error in any case. Since the certificate authencity is not enforced on OpenSSL-library level but inside postfix later, the error is not enforced. The error messages generated stay however in the error queue. - For blocking sockets, the SSL_accept()/connect() calls return "success", so the error-queue is never checked. - With BIO-pairs, the error queue is checked to find out, whether the function has just to be called again to continue the handshake, so the error messages are found and the connection is shut down due to the error condition. - Submitted bug report to Bodo Moeller. Bug fix is checked into the OpenSSL CVS archive: if the error is ignored during the handshake, clear the error-queue. * The next release of OpenSSL will behave consistently. - This leaves open the question, why the RSA-padding error is issued in the first place. Sent a query to the OpenSSL-* mailing lists. 2000/05/26 - A second site experiencing this problem pops up. -> Issued a warning to the postfix_tls mailing list. 2000/05/24 - Contacted Damien Miller . He did not change his TLS setup in the last time. He is running Postfix/TLS-0.6.6. - Contacted Bodo Moeller , the author of the BIO-pair part of OpenSSL for some debugging hints. Received several worthful remarks on what to look for. - Checked byte-for-byte the data fed into the OpenSSL-library. It does not differ between 0.6.9 and 0.6.10, so my handling seems to be actually correct. 2000/05/23 - A communication error occurs when talking to mail.mindrot.org: SSL_accept error -1 10264:error:0407006A:rsa routines:RSA_padding_check_PKCS1_type_1:block type is not 01:rsa_pk1.c:100: 10264:error:04067072:rsa routines:RSA_EAY_PUBLIC_DECRYPT:padding check failed:rsa_eay.c:396: 10264:error:0D079006:asn1 encoding routines:ASN1_verify:bad get asn1 object call:a_verify.c:109: - The error occurs both in client and server mode. 0.6.9 does not show this problem. - Tried to connect with several other sites, all connections are fine, this includes sendmail and qmail peers; hence decided to not recall 0.6.10. 2000/05/23 == Released 0.6.10 == 2000/05/23 - Sent a note to openssl-dev@openssl.org about the behaviour of SSL_free() and BIO_free(), hoping for some clarification whether my way of doing it is the recommended way. - Run the software in production mode on my own servers... - Finished writing the in-source documentation. - Updated sample-tls.cf and sample-smtp[d].cf to reflect the new timeout parameters. 2000/05/21 - Removed error messages produced by the now non-blocking behaviour of the TLS layer [apps_ssl_info_callback()]. 2000/05/20 - Took results home and tried to run it on my Linux-box: SEGV after successfully handling the SMTP session!! * It seems that the SSL_free() and BIO_free() functions interact. SSL_free() releases the underlying BIO and it will bomb out when it is then explicitely BIO_free()'ed again and vice versa. * It did not bomb out on HP-UX, but such things happen. I however want to know, why the example program does not fail... * With respect to the bevaviour as is, SSL_free(TLScontext->con); BIO_free(TLScontext->network_bio) and not touching TLScontext->internal_bio works. - Introduced special timeout values for the TLS negotiation stage, as the timeout values may change with protocol state (suggested by Wietse). - Started writing a full description of the BIO-pair concept and its special treatment into the pfixtls.c sourcecode. 2000/05/19 - Systematicly implemented a generalized layer handling: * do_tls_operation() is the generic handler for all SSL_*() input/output functions. It deals with the non-blocking behaviour of this functions, requiring appropriate retrys. * network_biopair_interop() handles the interaction between the socket/fd and the buffering BIO-pair. 2000/05/18 - Based on the example in openssl-0.9.5a/ssl/ssltest.c realized the first usage of BIO-pairs. (Can do server handshaking.) - Learned, that the BIO-pair has its own buffering that needs its own flushing. It is not enough to relay on the SSL_ERROR_WANT_READ/WRITE state information. 2000/05/17 == Released 0.6.9 == - Important: the seperator in the relay-fingerprints is now ':'!!! Don't forget to change your relay_clientcerts databases. 2000/05/16 - Changed pfixtls.c to only use the interface described in util/vstream.c for handling the VSTREAM. * Added vstream_context() macro to the VSTREAM-interface. - Introduce TLScontext to identify the connection instead of the file descriptor. Move all static data (SSL structure and information gathered about the connection) into the context. The TLScontext is allocated on TLS-start for a connection and saved with the VSTREAM, so several streams can be used at the same time. - Removed "pfixtls_setfd()" as it is no longer needed. - Changed the relay_clientcerts list from string_list_* to maps_* interface to allow usage of ":" in the list. THIS IS AN INCOMPATIBLE CHANGE!!!! - Updated documentation accordingly. 2000/05/12 == Re-released 0.6.8 == 2000/05/12 - Wietse announces snapshot-20000511 with an important bugfix. - Since upgrading from 20000507 to 20000511 is highly recommended, Postfix/TLS 0.6.8 is re-released for this snapshot (the patch applied cleanly, just the name of the toplevel directory has changed). 2000/05/11 == Released 0.6.8 == 2000/05/11 - Unlike expected I found some time to install the latest cyrus-sasl-1.5.21 and test some parts the integration. It does, well, work as advertised (and the advertisement in SASL_README is not too optimistic). - When checking all of the rejected patch-snippets for 0.6.6->0.6.7 I missed the parameter "smtpd_enforce_tls" (noted since I wanted to enforce TLS encryption while playing around with plaintext passwords) in the static CONFIG_BOOL_TABLE bool_table[] = {..} in smtpd/smtpd.c -> I will immediately release a corrected version 0.6.8. 2000/05/11 == Released 0.6.7 == 2000/05/11 - The latest sendmail.8.11.0.Beta1 includes STARTTLS support; it is available in source code and also uses OpenSSL. 2000/05/10 - After having it running at home (Linux) I also install it at work for the field test. - No time to install the SASL kit, so this part stays untested as of now. 2000/05/09 - Downloaded snaphot and apply the patchkit. - Straightened out the rejected parts of the patch. - Due to the new layering with timed_read() and timed_write() functions the integration of the TLS layer needed special adjustment. * When TLS is active, the timed_read() and timed_write() functions are replaced by the corresponding pfixtls_timed_read() and pfixtls_timed_write() functions. When the TLS functionality is stopped, the old functions are restored. * The names of the pfixtls_timed_*() functions are looking into the future, because they are working as before, the timeout functionality is not in, yet. 2000/05/08 - Wietse announces snapshot-20000507 with a lot of changes. Especially important: the I/O handling of the smtp-stream has been changed to a more layered technique that allows easier integration of the TLS layer. 2000/04/27 == Released 0.6.6 == 2000/04/27 - Fixed inconsistency between documentation and actual behaviour: peer certificate information was not logged at level 1 (found by Damien Miller ). * While at it: the logged information did not say whether the certificate data logged passed verification or not: fixed. (The information logged in the Received: header already contained that information.) - Backported dict_dbm.c from snapshot-20000309 with the updated dict_delete() behaviour (key not found is not considered fatal). Maintained dict_sdbm.c accordingly. 2000/04/18 == Released 0.6.5 == - Important: * New session cache mechanism SDBM. Please adapt your main.cf and delete any old ".db" session cache files manually. 2000/04/18 - I am using the SDBM session cache for a week right now and did not have any trouble, so I think its worth pushing it out. - I am not completely happy with the dict_del() behaviour of considering a not-found key fatal. It might happen when the smtp[d] processes would be allowed to delete themselves. They are not as of now, so I accept it for now but will reconsider it. - Updated documentation accordingly. 2000/04/17 - Received corrections for the HTML-docs from Ralf Hildebrandt . 2000/04/11 - Transfered SDBM from home (Linux-testbed :-) to work [found and fixed some small items when compiling on HP-UX]. Started running it under "real life" conditions. 2000/04/07 - Implemented "SDBM" Simple Database Management routines as also utilized in ModSSL. Of course, it requires reopening of the databases, so the routines are changed, that the _file_descriptors_ are left open, but the _in_memory_ database stuff (especially the cached data) is closed and reopened on access. This is what is really needed. The pagesize is increased from standard DBM compatibility to hold the session information. Additionally, this software is in the public domain, so no additional license problems arise. - The access goes through the dict_* interface, hence the locking is performed by myflock(). 2000/04/01 == Released 0.6.4 == 2000/04/01 - Updated to the new patchlevel of Postfix (19991231-pl06), some parts of the patch were rejected due to changes in smtpd. - Changed patch name with respect of today's release of OpenSSL-0.9.5a. The code remained unchanged. 2000/03/25-31 - The cached informations are not deleted by "tlsmgr" even though stored and retrieved by the smtp[d] processess. Strange. - Spend some large amount of time digging through the Berkeley DB documentation and code. * It claims that Berkeley DB is multi-process capable. Caveat: it takes the very complicated "transaction model", that I did not use until now. Hence the session cache does not work as is. * Even with transaction model, Berkeley DB requires re-opening of the databases to get rid of cached information. F*ck. - Finally, I give up on Berkeley DB for session caching. It will never work for us. Even if it would, it requires a large amount of helper files and it seems, that the transaction environment is somewhat fragile when it comes to some problem. I won't rely on it. 2000/03/28 == Released 0.6.3 == 2000/03/28 - As has been pointed out to me, the TLS information in the Received: header is not conform to RFC822. - The TLS protocol and peer CN information is now included in '()', so that it is a comment. 2000/03/21 == Released 0.6.2 == 2000/03/21 - I have been running DB based session caching with the changes for some more time now without problems. Am I really confident? No, not really. I remember the trouble I had with Berkeley DB and sendmail on HP-UX. I don't think I really trust it. - Realized single "smtp_tls_per_site" lookup. I cannot use the more or less comfortable "domain_list" lookups as before, since these do not return the value, just found or not :-(. Hence the lookup is realized with maps and exact lookup. I never tried regexp. But if I understand the docs correctly, it should be possible to use it here to realize wildcard lookups, if it would not have been disabled :-(. - Summary: * Session Cache will be cleaned at "postfix reload" or "postfix start" * New table "smtp_tls_per_site" * Gone: "smtp_tls_[use/enforce]_[recipients/sites]" 2000/03/16 - Changed pfixtls.c, so that it will only open Session Cache databases, that are already available. tlsmgr is responsible for creation. - Change tlsmgr.c, such that session cache databases will be removed before opening, so that fresh databases are used whenever postfix is restarted. This means, that session information is not kept over a postfix stop/start or reload sequence, but it also means, that issuing a postfix reload will clean the session cache. I don't use simple dict_open with O_TRUNC, because this would not help against database files, that are locked by hanging smtp[d] processes. If you think it will also solve the "hang" problem described for 2000/03/15: in a certain sense it can, since tlsmgr will be killed by the watchdog and new, fresh cache files are installed, but that is not more than an ugly hack. It must be solved in a clean manner. 2000/03/15 - Experienced some strange problem with Berkeley DB based session cache. The DB routines hang while trying to delete an entry. I did save the corresponding "hash:" file and could reproduce it (and walk through the endless loop with a debugger), but I didn't find the reason why. Since during "db->del" the database is exclusively locked all other processes hang however, so this is really bad!!!!!!!! 2000/03/12 == Released 0.6.1 == 2000/03/12 - Created tls_info_t structure to hold all information about the active TLS connection. Remove all global variables except those for the running client/server engines (those might be replaced with global variables in smtpd/smtp, though). - Added field "dNSName" to the structure (still unused). This will be used with X503v3 extensions. - Cleaned up TODO, since some items are now done... 2000/03/11 - Added missing #include to tlsmgr.c. (Worked without on HP-UX, showed up on Linux.) - Bug: removal of server side sessions from the cache in case of trouble failed, because uppercase hex was used instead of lowercase for the key. This does not affect removal of expired sessions by tlsmgr. - Stepped up to postfix-19991231-pl05. 2000/03/09 == Released 0.6.0 == - Important: * This release features an additional daemon, the "tlsmgr", please update your master.cf accordingly. * This release does not use the /var/spool/postfix/TLS* directories anymore. Remove them and re-install the original postfix-script. * Check the new/changed configuration parameters tls_random* and smtp[d]_tls_session_cache*. * This release will only work with OpenSSL >= 0.9.5!!!!! 2000/03/09 - Testcompilation of Postfix/TLS without -DSSL and the OpenSSL includes and libraries passed. - Worked through tlsmgr.c to remove unneeded header files. - Wrote documentation for tlsmgr.c. - Updated documentation on top of pfixtls.c. - Put (char *) casts into the myfree() calls, where necessary, to make the HP compiler happy. - Updated html PRNG documentation in Postfix/TLS. 2000/03/08 - Finished first version of "tlsmgr". Does run through session cache databases and detects and deletes (*) old sessions. * Had to realize SYNC_UPDATES for the dict_db_delete() function and patch the flag handling within the function. Changes sent to Wietse. - Restored qmgr to its original state. - Extended pfixtls.c to need an additional "needs_095_or_later()" function when compiled with an older version of postfix. - The session cache is now enabled, when a database filename is given. smtp[d]_tls_use_session_cache configuration parameters removed, updated documenation accordingly. - Moved the PRNG handling to tlsmgr, applying the new model. tlsmgr will query external sources at startup and will then feed a PRNG exchange file with random data in intervals of configurable (but random driven) length. If running outside chroot, tlsmgr can query the entropy source (e.g. EGD or /dev/urandom) again and so increase entropy with time. If the entropy sources don't limit access, the tlsmgr can run with "postfix" privileges. Mine does. -> master.cf became a new entry. - tlsmgr is realized as a trigger server and has the "fifo" entry. Actually, it does not take any input. One could utilize it to feed back some entropy from running smtp[d] processes, but I think this would overload the issue. - I will release a 0.6.0 pre-version as is. tlsmgr still lacks the detailed information in the header and the interface description in pfixtls.c probably is also not longer up do date. 2000/03/07 - Since defective session data can cause SEGFAULTs, it is now armored by a leading structure that does contain a session cache version and the postfix library version before the timestamp. If a session does not match exactly the version numbers, it is immediately discarded and deleted to avoid harm. - Removed the seperate storage of the peer's certificate verify_result, so starting from this moment, Postfix/TLS will only work safely with OpenSSL >= 0.9.5!!! - Ported server side session cache routines to the client side; works. - Analyzed structure of "qmgr" to understand consequences for the planned "tlsmgr" daemon. Transferred the sceleton. - Received word from sendmail, a (at least preliminary) TLS enabled test address is "bounce@esmtp.org". 2000/03/06 - Wietse supplied a change to the dict/dict_db mechanism to allow for synchronous updates. Session cache updates for the server side seem to work now, removal of old sessions (when called from the client) integrated. 2000/03/05 - Got the database style session cache to run for the server side (at least partial). The removal of old sessions is not yet realized. [There are several man pages for OpenSSL as of 0.9.5, but the i2d etc interfaces are not belong them, so I had to study the source code instead.] * What is not working by now is the synchronization of the memory database to disk. It only is synchronized automatically upon close. It would be necessary to sync after each update or delete, but this is not implemented in Wietse's dict library. I will post an according proposal. 2000/03/04 - Wietse posts a patch to select "EHLO" negotiation even if ESMTP is not recognized from the 220 greeting. Activating this flag will however break compatibility with mailers, that simply close the connection upon EHLO. I don't know how the large the number of these broken mailers is, but activating "smtp_always_send_ehlo" is a tradeoff. - Integrated Wietse's patch into Postfix/TLS. 2000/03/03 - Received update from Matti Aarnio (Zmailer) is now for some time able to do server _and_ client side TLS. Updated documenation accordingly. When testing, Postfix client to Zmailer server failed, because Zmailer announces with "ESMTP+IDENT" and Postfix does not recognize the ESMTP token (must be seperate), so only HELO is used and STARTTLS is not offered by the Zmailer server. Informed Matti accordingly, will wait until the problem is resolved before actually publishing the update. - Enhanced the documentation by listing automatic reply services at which interoperability can be tested. 2000/03/02 - Went through the Postfix source to check out the database routines. It should be possible to move session caching from directory/file- based to database. Since DBM only allows blocks (key+contents) of 1024 bytes and a session is larger, only Berkeley DB can be used. Put some first bits into Postfix/TLS. 2000/02/29 == Released 0.5.5 == 2000/02/29 - OpenSSL 0.9.5 has been released. Since I want to promote 0.9.5, as it contains several bugfixes and enhancements, I release a new version of Postfix/TLS. My personal highlights: * The bug with Win32 Netscape not commencing after certificate storage unlocking should be fixed. (I will leave the not in however, as long as I have not positively checked it myself. Reproducibility...) * The bug, that the certificate verifiation result is not stored in the session cache (discovered for Postfix/TLS 0.4.4) is fixed. I will leave the Postfix/TLS workaround in as long as it will run with older versions of OpenSSL. * The OpenSSL commandline tools like "openssl gendh" now support EGD, so that the examples for generating the DH parameters now will really work with high quality random data :-) * The support of 56bit ciphers has lost its importance since 128bit versions of Netscape etc are now easily available... - This version does not feature source code changes but updated documenation when compared with 0.5.4: * List examples on how to generate good entropy for the PRNG seed in /etc/postfix/random_file. - Update the TODO document with respect to the discussion about session caching and other security items. This document is a very short summary, for the full discussion check the mail archive at http://www.aet.tu-cottbus.de/mailman/listinfo/postfix_tls/ 2000/02/26-28 - Wietse considers including Postfix/TLS into the main release. A discussion about security relevant features, especially the session cache inside the chroot jail takes place. The discussion will definetely lead to some changes; I have however not decided on the first step, yet :-) 2000/02/21 == RELEASED 0.5.4 == - Important: Another directoy is created in /var/spool/postfix, so don't forget to install the new versions of conf/postfix-script-*sgid. 2000/02/21 - Finished the seed-exchange architecture by saving the random seed at exit of smtp and smtpd. - Wrote documentation for the PRNG handling to the documentation. - Tested on HP-UX (with a current OpenSSL-pre-0.9.5 snapshot and 0.9.4) and on SuSE-Linux (with 0.9.4). * THIS VERSION WILL STILL RUN WITH OPENSSL-0.9.4, but it will also run with OpenSSL-0.9.5. Older versions of Postfix/TLS will not, because the PRNG is not seeded! 2000/02/19 - Start to implement my own model of collecting entropy. All smtp and smtpd processes will record some items (mainly the time of actions) to add some entropy into the PRNG. The state is saved and used to re-seed by the smtp and smtpd processes, so that entropy adds up into the pool. The seeding by external file is additionally kept in order to be able to inject additional entropy. 2000/02/18 - Included routines to add random seed from a configurable file "rand_file_name". I don't want to retrieve the entropy from a real random system source, because the amount of entropy that can be collected is limited. We might hence stall. Let's think about this problem. - The SSL_CTX_load_verify_locations() has been fixed in the latest OpenSSL snapshot. 2000/02/17 - Tracked down the SSL_CTX_load_verify_locations() problem in the OpenSSL library. If more than one CA-certificate is loaded, a bogus return value 0 is created, because the count of certs is checked to be "1" instead of allowing ">=1". Reported to openssl-dev. 2000/02/16 - Downloaded the latest openssl-SNAPSHOT-20000215 and installed it on my development machine, then recompiled Postfix/TLS and try to run it. * Failure: SSL_CTX_load_verify_locations() fails on reading the CAfile with return value 0, but no actual error is displayed. If the return value is not checked, the CA-certificates work, so that they are loaded and the error indicator seems to be bogus. Reported to openssl-dev mailing list. * Failure: OpenSSL has become picky about correct seeding of the PRNG Pseudo Random Number Generator. Installed some "testseed" that is actually not random, but then Postfix/TLS starts to work again. We will need some good random seed setup, probably reading from either /dev/random (if available) or from EGD. Found out during the experiments, that EGD is not that simple to use as described in some of my Postfix/TLS docs. Must be upgraded. Asked in the openssl-dev mailing list about the recommended amount of random data needed for seeding the PRNG. Ulf Moeller recommends a minimum of 128bit. 2000/02/14 == Released 0.5.3 == 2000/02/14 - OpenSSL 0.9.5 is to be released within the next hours/days. Since I intend to use some of its new features soon, I will re-release 0.5.2 as the last version that will run with 0.9.4 but for the latest postfix patchlevel. - No functional changes. - Updated patch for postfix-19991231-pl04. 2000/01/28 == Released 0.5.2 == 2000/01/28 - Stepped up the next postfix patchlevel postfix-19991231-pl03. No functional changes. 2000/01/03 == Released 0.5.1 == 2000/01/03 - Bug fixed: Don't specify a default value for "smtpd_tls_dcert_file", assuming that typically a DSA certificate is not used. Otherwise smtpd will try to read it on startup and the TLS engine won't start since it is not found. I didn't note this bug before today, because I could not install this release in a larger scale on my own servers due to a network failure of our campus backbone lastring from Dec 31 until today. - Stepped up to the just released postfix-19991231-pl01. 2000/01/01 == Released 0.5.0 == 2000/01/01 - Upgraded to the new postfix release 19991231. 1999/12/30 - Enabled support for DSA certificate and key for the server side. One can have both at the same time, the selected cipher decides which one is used. OpenSSL clients (like Postfix/TLS) will prefer the RSA cipher suites, if not especially changed in the cipher selection list. Netscape will only use the RSA cert. - The client side can only have one certificate. There is a way out by using a callback function, that will receive the list of acceptable CAs and then do some clever selection: SSL_CTX_set_client_cert_cb(). I will however have to figure out, how it has to be prepared, it seems, that there is no example available. - I have been able to successfully generate a DSA CA and certificates for some Postfix hosts and to do authentication and relaying as expected. So now I have to document how it is done in a practical manner... - Moved up prerelease 0.5.0pre02 to the download site. 1999/12/28 - Moved up to SNAPSHOT-19991227. - Don't forget to check the return value when calling SSL_CTX_set_cipherlist(). - Add code to load DH-parameters from disk. - Add configuration information for the new functionality: DH paramter support, possibility to influence the cipherlist. - Moved up prerelease 0.5.0pre01 to the download site. 1999/12/25 - Found some minutes to relax from the christmas business. - Applied the 0.4.7 patch to SNAPSHOT-19991223 and included the new changes of 1999/12/19. Once the new stable release of postfix is out, this minimum state will be the new Postfix/TLS patch: the new functionality will not influence stability, so it can stay in even if still unfinished. 1999/12/23 - Wietse announces SNAPSHOT-19991223: if no severe bugs are found, it will be promoted as next stable release soon. Good to have kept everything from yesterday. 1999/12/22 - Got a query from a Postfix/TLS user: the patch does not apply cleanly to SNAPSHOT-19991216 and he somehow messed up to integrate the rejected parts (it later turned out he just forgot on reject). Applied the patch myself and generated a diff, sent it to the user and of course kept a copy for myself, since I will have to apply it myself eventually once the next "stable" release of postfix is out. 1999/12/19 - Began modifications for 0.5.x: * Added configuration variables for specifying the cipherlist to be used smtpd_tls_cipherlist and smtp_tls_cipherlist. For the format, there is some (however sparse) documentation in the openssl package. * Call SSL_CTX_set_cipherlist() with these data. * Added default temporary DH parameters to pfixtls.c (only server side is necessary) and configuration variables to specify user generated parameters; they are however not used, yet. The default parameters were generated using the presumably good /dev/random source. 1999/12/13 == Released 0.4.7 == 1999/12/13 - Addendum to the last change: do also remove sessions, that could _not_ be reused. - Updated configuration information: * As of OpenSSL 0.9.4, certificate chain verification is not sufficient, since the certificate purpose is not checked, so I recommend to add all intermediate CAs the the list of CAs and stay with a verification depth of 1. Work is in progress for 0.9.5. - Stepped up to the just released new patchlevel postfix-19990906-pl09. 1999/12/10 == Released 0.4.6 == 1999/12/10 - Realized changes implied below: Removed SSL_CTX_add_session() in the client startup; remove session on stop with SSL_SESSION_free(). - In the morning there is a mail on the list, that Postfix might be crashed with a single "\" on the "CC:" line. Hence, we should expect a new patchlevel soon. Release the actual change anyway. 1999/12/09 - Read in the "openssl-users" mailing list, that SSL_CTX_add_session() is only intended for servers. On the client side, SSL_set_session() is sufficient. Additionally, the session should be explicitely freed, since SSL_set_session() will increment the usage count for the session. Explained by Bodo Moeller. 1999/12/xx - Had a discussion (by email) with Bodo Moeller about DH/DSS. It seems I understand better now (after the discussion) how it works :-). Implementing it should not be too difficult but might take some more hours. Mentally scheduled it for Version "0.5.0" whenever this might be (rough guess: christmas vacation). Decided to hence not discuss this topic in the docs, since it might change in the near future anyway. 1999/11/23 - Discussion with rch@writeme.com (Richard) about implementing DH ciphers and DSA keys and certificates on the Postfix/TLS list: It does not work as of now. 1999/11/15 == Released 0.4.5 == 1999/11/15 - Applied patch to postfix-19990906-pl07 without problems. Well, let's release new version of Postfix/TLS, so that we look up to date. - Add the "DO NOT EDIT THIS FILE" to conf/sample-tls.cf. 1999/11/08 - Applied patch to the fresh release of postfix-19990906-pl06 without problems. Nothing else, so no new release of Postfix/TLS. 1999/11/07 == Released 0.4.4 == 1999/11/07 - Played around some more with the X509_verify_cert() function: when saving a session, neither the verify_result is saved nor the certificate chain necessary to re-verify. So there were two possibilities left: do a full renegotiation negating the benefit of session caching or - save the verify_result into to the session cache file and set the value when rereading from disk. This way the positive result of session caching is kept. - Make sure, the verify_result value is propagated as pfixtls_peer_verified and used where needed. - After experiencing some failures at TLS connection setup, the SSL_sessions are now freed again when closing. It seems, something is left over in the session structures, even though SSL_clear() is called. 1999/11/06 - When not asking for a client certificate, the "Received:" header will show the protocol and cipher, but silently omit the client CN (because they where not supplied). Noted by Craig Sanders . The same holds, if a certificate is asked for, but none supplied. Now, in any case an appropriate information is added in the "Received:" header. - Added a hint to remove sessions from the cache during testing, since old information may still be in the cache. Also proposed by Craig Sanders . - While at it: client CN and issuer CN are printed, but the verification state is not, so that the trust value of this data is not known. * Added (verify OK/not verified) to the Received: header. * Obtained information using the SSL_get_verify_result(SSL *con) call. * Learned, that the state is not saved in the session information, so that a recalled old session will always return "OK" even if the certificate failed the verification! Call it a bug in OpenSSL. Still investigating on a good way to work around this problem. - Fixed a bug in the syslog entries: The client CN is logged, but the issuer CN is not, because of a missing "%s" in the format string. 1999/11/03 == Released 0.4.3 == 1999/11/03 - Added some hints about security to the html documentation. - Tested the changes made two weeks ago at home in the large university setup. I was to a conference in between and didn't want to release the new version without having done some more tests. 1999/10/17 - Added another half a ton of comments (this time for the client side), yielding one ton alltogether... 1999/10/16 - Rearranged some of the TLS-engine initialization to improve readability. - Do not "free" the SSL connection, when it is not really necessary. Do only reset information about the TLS connection, when there was one. This is the better way instead of the quick fix applied for 0.4.2. - Added half a ton of comments to the TLS code (server side) to document what is done when and why, since there is no real documentation about the OpenSSL library. 1999/10/11 == Released 0.4.2 == 1999/10/11 - Fixed a severe bug introduced in 0.4.0: smtpd and smtp tried to flush old session from the session cache even when TLS was not enabled. Since no SSL-context was allocated, smtp would segfault on connection close. 1999/10/10 == Released 0.4.1 == 1999/10/10 - Added a long description of the session cache handling to the top of global/pfixtls.c. - There is a race condition when cleaning up the session cache in qmgr, that might lead to lost sessions in client mode. The worst consequence is an additional session negotiation, so we can live with it as of now. Bug described in qmgr/qmgr_tls.c. - Implemented immediate removal of session cache files with expired sessions when these are called. No need to first load and then discard them. - Implemented the requirement from RFC2246 to remove sessions, when connection failures occure (well actually, when TLS layer failures occur, but I cannot seperate this from another) for the server side. the client side is under work. 1999/10/09 - Set an absolut maximum length of 32 for the IDs used for session caching. This matches the default in OpenSSL, but I don´t want to see surprises when somebody sometimes will run into a longer session id. 1999/10/05 == Released 0.4.0 == - The new disk based session cache is a major step, so the minor release number is pushed to 0.4. - By now I think all necessary bells and whistles are in the code. What is left is a big code cleanup and some more testing before calling this patchkit "1.0.0". - Initiated Mailing List at http://www.aet.tu-cottbus.de/mailman/listinfo/postfix_tls 1999/10/05 - Some code cleanup. - Added new options to the documentation and the hint to update "postfix-script", because otherwise qmgr might fail! 1999/10/03 - Realized disc based session caching also for the Postfix/TLS client. Must go to real world testing now between hosts. And, of course, tune up the documentation, because users will have to install a new postfix-script, too. 1999/10/02 - The old sessions must be removed once they have timed out, so a process is needed that will scan through the list of old sessions and remove once they have expired. Lucky me: this is what qmgr usually does with deferred messages, so qmgr is extended only a little bit and will now also clean up the old sessions from the cache directory. And hey: it is good to see how easily this thing can be extended and functions can easily be reused. Postfix is an excellent peace of software engineering and there is no line of C++ or other "object oriented modern junk" in it. It should be recommended as an example to computer sience students. 1999/09/28 - I cannot use the mod_ssl way for session caching and I don´t want to spend an extra "gcache" daemon as ApacheSSL does. So I follow Wietse´s idea realized for his mail queues and create hash level based subdirectory structures. The good thing: I can cannibalize the mail_queue code. The bad thing: there is a path length of 100 chars fix coded in Wietse´s routines. It does hold for 32byte session ideas. Status: can save sessions to disk and recall them (server side). 1999/09/26 - Created new call backs for external session caching for the server side. In a first step, they can print out the session ids for the newly created session and when recalling a session. As the OpenSSL documentation on this is pretty sparse, Ben Laurie´s ApacheSSL code is very helpful, Ralph Engelschall´s Mod_SSL code for session caching is far more complicated. 1999/09/23 == Released 0.3.10 == 1999/09/23 - Debugging for 0.3.8/0.3.9 would have been so much easier, if the error messages put onto the error message stack from the OpenSSL library would have been printed out. The error was clearly stated from the library, I just didn't print it. Added pfixtls_print_errors() calls where missing after calls to the OpenSSL library. Sometimes I feel so old... - Used opportunity to upgrade to the latest postfix patchlevel 05: postfix-19990906-pl05. 1999/09/19 == Released 0.3.9 == 1999/09/19 - Added a "smtp_no_tls_sites" table to allow people to enable TLS negotiation globally and only omit it on a per site basis. 1999/09/18 - Finally found the bug described for 0.3.8: In the server setup, the SSL_CTX_set_session_id_context() call was missing. To find this, I had to trace through the OpenSSL library and when I finally found it in ssl/ssl_sess.c, there was an appropriate comment about this. I however have to find out why I didn´t receive the appropriate error message... - This bug was hidden during the first developing stages, as the shutdown sequence was not working correct, so the session was not cached. 1999/09/17 == Released 0.3.8 == 1999/09/17 - Something is strange with the session caching in smtpd server mode with Netscape 4.61 client. The first connection is fine, the next one hangs after the server fails with errors while reading the SSLv3 client hello C. (Found by Michael Stroeder ) Reproducable with OpenSSL 0.9.3a, 0.9.4 and SNAPSHOT 19990915, so the problem seems to be persistent. I will try to figure out the problem myself before reporting it to the developers. If I don't find it, maybe they do :-) Workaround: the cached session is removed after connection is closed. This will impose some time penalty on the negotiation. As the caching is local in the smtp processes and they time out anyway, the penalty should not be significant. The problem does not occure with Postfix/TLS clients. 1999/09/13 == Released 0.3.7 == 1999/09/13 - Ran tests, seems no further conflicts between Wietse's changes and my extensions. 1999/09/09 - Applied the patchkit 0.3.6 to postfix-19990906-pl02 and worked out the rejected part of the patch. From this point of view the patch is included. Now everything has to be retested. 1999/09/09 == Released 0.3.6 == 1999/09/09 - Added a missing ´#ifdef HAS_SSL #endif´ in smtp_connect.c. Noted by Jeff Johnson . - HINT: On 1999/09/06 a new "stable" version of postfix was released. Future Postfix/TLS enhancements will be against this new version 19990906. 1999/08/25 == Released 0.3.5 == 1999/08/25 - Added Wietse's patch for postfix-19990601 to prevent crashing smtpd when VRFY is called without setting the sender with "MAIL FROM:" first. 1999/08/13 - Small changes to global/pfixtls.[ch]: Since we also support client STARTLS, we check the peers certificate, which may also be a "server" certificate (not just client). Hence I renamed "*ccert*" to "*peer*". - global/pfixtls.c: add some "const" to "char *" for OpenSSL library calls, to make gcc happy. - Extended comments in pfixtls.[ch] to better match Wietse's style. 1999/08/12 == Released 0.3.4 == 1999/08/12 - Enabled workarounds for known bugs in SSL-engines. - Tested with OpenSSL 0.9.4. - Windows95/NT: Problem with Netscape hanging on first connection when the client certificate database has to be unlocked cannot be reproduced anymore. I am happy, but I am also not sure what caused the problem to go away and I cannot figure out the security settings manually from the files... 1999/08/11 - Corrected loglevel handling: At some points smtpd_tls_loglevel was used instead of smtp_tls_loglevel (only noted at loglevels >= 2). 1999/08/09 == Released 0.3.3 == 1999/08/09 - Removed SSL_CTX_set_quiet_shutdown() as it does prevent the shutdown from actually being performed. In order to remove the annoying "SSL3 alert write:warning:close notify" it is now explicitly handled in apps_ssl_info_callback(). Bug found by Bodo Moeller . 1999/08/06 == Released 0.3.2 == 1999/08/06 - Add option "smtp_tls_note_starttls_offer" to collect information about hosts, that offered the STARTTLS feature without using it. - Shut up smtpd. Only print information about relaying based on certs when msg_verbose is true. 1999/07/20 - Added missing "const" in pfixtls.h (found by Juergen Scheiderer ). HP-UX ANSI-C didn't complain. 1999/07/08 == Released 0.3.1 == 1999/07/08 - New config variable "smtpd_tls_received_header". When "true", the protocol and cipher data as well as subject and issuer CN of the client certificate are included into the "Received:" header. 1999/07/07 - "starting TLS engine" message will only be printed when loglevel >=2 to reduce unnecessary noise in the log files. - Added code to fetch the protocol (e.g. TLSv1) and the cipher used (by name and bits). Information is printed to the logfile. 1999/07/01 == Released 0.3.0 == 1999/07/01 - (Client mode) Bug fix: Don't try to use STARTTLS if it is not offered. The server we are connected to might not understand it and respond with a "500 command not understood", causing the email to bounce back, even when the lack of STARTTLS is just a temporary problem. - Updated documentation for the new per recipient/site TLS decisions. 1999/06/30 - Client mode: Added variables and routines to decide "per recipient" or "per host/site" whether to use/enforce TLS or not. 1999/06/18 == Released 0.2.8 == 1999/06/18 - In client mode the "use_tls" and "enforce_tls" internal variables were not initialized correctly, such that the client could try to use the STARTTLS negotiation even if not wanted. This error was introduced in 0.2.7. Noted by "Cerebus" . 1999/06/08 == Released 0.2.7 == 1999/06/08 - Studied discussions in the IETF-apps-TLS mailing list: MS Exchange seems to offer STARTTLS even if not configured. Added this info to the documentation. - Updated Documentation regarding the changes made. 1999/06/03 - The subject-CommonName (CN) of the server certificate is extracted when connecting to a TLS server. - In "smtp_*_tls" mode, this subject-CommonName is matched against the hostname of the server. In "enforce" mode, the connection is droppend when the certified server name and the real hostname differ. - Added missing dependencies in smtp/Makefile.in (missing pfixtls.h since 0.2.0). 1999/06/02 == Released 0.2.6 == 1999/06/02 - Adapted patchkit to postfix-19990601. 1999/06/01 == Released 0.2.5 == 1999/06/01 - Updated OpenSSL API to 0.9.3a -> position of include files has changed from to . No functional changes. - pkcs12 utility is now part of OpenSSL -> changed documentation accordingly. 1999/05/20 == Released 0.2.4 == 1999/05/20 - Updated postfix base 19990317 from pl04 to pl05. 1999/05/14 == Released 0.2.3 == 1999/05/14 - Fixed a bug in pfixtls_stop_*(): there was a ";" to much directly after "if (con);". This check is only done as a safety measure: When SSL is not started you should not stop it. This case could however only happen when the code in smtp[d] would be wrong, so it should never be necessary. (Bug found by Uwe Ohse ) 1999/05/11 == Released 0.2.2 == 1999/05/11 - Matti Aarnio: Reworked pfixtls_dump() to use fewer strcpy and strcat calls. - Added information about Matti Aarnio (author/maintainer of ZMailer) working on RFC2487 for ZMailer. 1999/05/04 == Released 0.2.1 == 1999/05/04 - Stuffed up the documenation to reflect the actual status. No change in functionality. 1999/04/30 == Released 0.2.0 == 1999/04/30 - Adjusted the changes in smtp*.c to Wietse's indentation style. - Sorry, the documentation about the client side has by now to be taken from sample-tls.conf. The documenation has to be rearranged in a larger scale. 1999/04/29 - Finished client support for STARTTLS in smtp; some testing done. - Fixed a race condition in smtpd: When in PIPELINE mode, the connection was switched back from SSL to normal mode before the buffers were flashed. - Adjusted the code in pfixtls.[ch] and additions in smtpd*.c to Wietse's indentation style. 1999/04/28 - Incorporated skeleton of STARTTLS support into smtp. - Introduced variables to control client STARTTLS to configuration. 1999/04/15 == Released 0.1.5 == 1999/04/15 - Adjusted pfixtls.diff to postfix-19990317-pl04. 1999/04/14 - Ported from OpenSSL the BIO_callback functions to dump out the negotiation and transmission for debugging purposes. The functions are triggered by the the new loglevels 3 and 4. - Call SSL_free() to get rid of the SSL connection structure not used anymore. 1999/04/13 == Released 0.1.4 == 1999/04/13 - Based on a hint in the openssl-users list added an SSL_set_accept_state() before the actual SSL_accept(). I don't really understand why, but the documentation of SSL is a bit short anyway. 1999/04/11 - Some more comments on certificates in the documentation. 1999/04/10 - Moved initialization of the pfixtls_server_engine to the pre_jail_init() section of smtpd, so that it is called with root privileges to read the key and cert information. The secret key of the server can now be protected by "chown root secretkey.pem; chmod 400 secretkey.pem". Additionally, this makes it possible to run smtpd in chroot jail, even though I didn't test that, yet. All information is read at smtpd startup time except the CAcerts in tls_CApath, which are checked at runtime. I have to look into that. - Updated documentation accordingly. - Rewrote the documentation with regard to the certificate setup and explaining the different types of certificates. 1999/04/09 - Introduced pfixtls_print_errors() which imitates BIO_print_errors() (the typical way to print error information in OpenSSL) but writes to syslog instead of a file handle. Hence we can get more informative error information. 1999/04/08 == Released 0.1.3 == 1999/04/08 - Stuffed up the documentation by reworking the references. - Added contributed script for automatic addition of fingerprints. - Added ACKNOWLEDGEMENTS file 1999/04/06 == Released 0.1.2 == 1999/04/06 - Portability: removed call of "snprintf()", as it is not available on some (older) UNIX versions (in this case Solaris 2.5). - Removed calls to "select()" when in TLS mode: Even though no new bytes arrive, there might be bytes left in the SSL buffer -> possible hang. 1999/03/30 == Released 0.1.1 == 1999/03/30 - Added disclaimer about export restrictions. - Fixed a bug in util/match_ops.c: When using dictionary lookup the compare was case sensitive by accident. Effect: Fingerprint matching did not work with databases, only for plain file. Bug report submitted to postfix author. 1999/03/29 == Released first version 0.1.0 ==