Change to mitigate branch prediction attacks

[thirdparty/openssl.git] / CHANGES

 OpenSSL CHANGES
 _______________

 Changes between 0.9.8e and 0.9.8f  [xx XXX xxxx]

  *) Mitigate branch prediction attacks, which can be practical if a
     single processor is shared, allowing a spy process to extract
     information.  For detailed background information, see
     http://eprint.iacr.org/2007/039 (O. Aciicmez, S. Gueron,
     J.-P. Seifert, "New Branch Prediction Vulnerabilities in OpenSSL
     and Necessary Software Countermeasures").  The core of the change
     are new versions BN_div_no_branch() and
     BN_mod_inverse_no_branch() of BN_div() and BN_mod_inverse(),
     respectively, which are slower, but avoid the security-relevant
     conditional branches.  These are automatically called by BN_div()
     and BN_mod_inverse() if the flag BN_FLG_CONSTTIME is set for the
     modulus.  Also, BN_is_bit_set() has been changed to remove a
     conditional branch.

     BN_FLG_CONSTTIME is the new name for the previous
     BN_FLG_EXP_CONSTTIME flag, since it now affects more than just
     modular exponentiation.  (Since OpenSSL 0.9.7h, setting this flag
     in the exponent causes BN_mod_exp_mont() to use the alternative
     implementation in BN_mod_exp_mont_consttime().)  The old name
     remains as a deprecated alias.

     Similary, RSA_FLAG_NO_EXP_CONSTTIME is replaced by a more general
     RSA_FLAG_NO_CONSTTIME flag since the RSA implementation now uses
     constant-time implementations for more than just exponentiation.
     Here too the old name is kept as a deprecated alias.

     BN_BLINDING_new() will now use BN_dup() for the modulus so that
     the BN_BLINDING structure gets an independent copy of the
     modulus.  This means that the previous "BIGNUM *m" argument to
     BN_BLINDING_new() and to BN_BLINDING_create_param() now
     essentially becomes "const BIGNUM *m", although we can't actually
     change this in the header file before 0.9.9.  It allows
     RSA_setup_blinding() to use BN_with_flags() on the modulus to
     enable BN_FLG_CONSTTIME.

     [Matthew D Wood (Intel Corp)]

  *) In the SSL/TLS server implementation, be strict about session ID
     context matching (which matters if an application uses a single
     external cache for different purposes).  Previously,
     out-of-context reuse was forbidden only if SSL_VERIFY_PEER was
     set.  This did ensure strict client verification, but meant that,
     with applications using a single external cache for quite
     different requirements, clients could circumvent ciphersuite
     restrictions for a given session ID context by starting a session
     in a different context.
     [Bodo Moeller]

  *) Include "!eNULL" in SSL_DEFAULT_CIPHER_LIST to make sure that
     a ciphersuite string such as "DEFAULT:RSA" cannot enable
     authentication-only ciphersuites.
     [Bodo Moeller]

 Changes between 0.9.8d and 0.9.8e  [23 Feb 2007]

  *) Since AES128 and AES256 (and similarly Camellia128 and
     Camellia256) share a single mask bit in the logic of
     ssl/ssl_ciph.c, the code for masking out disabled ciphers needs a
     kludge to work properly if AES128 is available and AES256 isn't
     (or if Camellia128 is available and Camellia256 isn't).
     [Victor Duchovni]

  *) Fix the BIT STRING encoding generated by crypto/ec/ec_asn1.c
     (within i2d_ECPrivateKey, i2d_ECPKParameters, i2d_ECParameters):
     When a point or a seed is encoded in a BIT STRING, we need to
     prevent the removal of trailing zero bits to get the proper DER
     encoding.  (By default, crypto/asn1/a_bitstr.c assumes the case
     of a NamedBitList, for which trailing 0 bits need to be removed.)
     [Bodo Moeller]

  *) Have SSL/TLS server implementation tolerate "mismatched" record
     protocol version while receiving ClientHello even if the
     ClientHello is fragmented.  (The server can't insist on the
     particular protocol version it has chosen before the ServerHello
     message has informed the client about his choice.)
     [Bodo Moeller]

  *) Add RFC 3779 support.
     [Rob Austein for ARIN, Ben Laurie]

  *) Load error codes if they are not already present instead of using a
     static variable. This allows them to be cleanly unloaded and reloaded.
     Improve header file function name parsing.
     [Steve Henson]

  *) extend SMTP and IMAP protocol emulation in s_client to use EHLO
     or CAPABILITY handshake as required by RFCs.
     [Goetz Babin-Ebell]

 Changes between 0.9.8c and 0.9.8d  [28 Sep 2006]

  *) Introduce limits to prevent malicious keys being able to
     cause a denial of service.  (CVE-2006-2940)
     [Steve Henson, Bodo Moeller]

  *) Fix ASN.1 parsing of certain invalid structures that can result
     in a denial of service.  (CVE-2006-2937)  [Steve Henson]

  *) Fix buffer overflow in SSL_get_shared_ciphers() function. 
     (CVE-2006-3738) [Tavis Ormandy and Will Drewry, Google Security Team]

  *) Fix SSL client code which could crash if connecting to a
     malicious SSLv2 server.  (CVE-2006-4343)
     [Tavis Ormandy and Will Drewry, Google Security Team]

  *) Since 0.9.8b, ciphersuite strings naming explicit ciphersuites
     match only those.  Before that, "AES256-SHA" would be interpreted
     as a pattern and match "AES128-SHA" too (since AES128-SHA got
     the same strength classification in 0.9.7h) as we currently only
     have a single AES bit in the ciphersuite description bitmap.
     That change, however, also applied to ciphersuite strings such as
     "RC4-MD5" that intentionally matched multiple ciphersuites --
     namely, SSL 2.0 ciphersuites in addition to the more common ones
     from SSL 3.0/TLS 1.0.

     So we change the selection algorithm again: Naming an explicit
     ciphersuite selects this one ciphersuite, and any other similar
     ciphersuite (same bitmap) from *other* protocol versions.
     Thus, "RC4-MD5" again will properly select both the SSL 2.0
     ciphersuite and the SSL 3.0/TLS 1.0 ciphersuite.

     Since SSL 2.0 does not have any ciphersuites for which the
     128/256 bit distinction would be relevant, this works for now.
     The proper fix will be to use different bits for AES128 and
     AES256, which would have avoided the problems from the beginning;
     however, bits are scarce, so we can only do this in a new release
     (not just a patchlevel) when we can change the SSL_CIPHER
     definition to split the single 'unsigned long mask' bitmap into
     multiple values to extend the available space.

     [Bodo Moeller]

 Changes between 0.9.8b and 0.9.8c  [05 Sep 2006]

  *) Avoid PKCS #1 v1.5 signature attack discovered by Daniel Bleichenbacher
     (CVE-2006-4339)  [Ben Laurie and Google Security Team]

  *) Add AES IGE and biIGE modes.
     [Ben Laurie]

  *) Change the Unix randomness entropy gathering to use poll() when
     possible instead of select(), since the latter has some
     undesirable limitations.
     [Darryl Miles via Richard Levitte and Bodo Moeller]

  *) Disable "ECCdraft" ciphersuites more thoroughly.  Now special
     treatment in ssl/ssl_ciph.s makes sure that these ciphersuites
     cannot be implicitly activated as part of, e.g., the "AES" alias.
     However, please upgrade to OpenSSL 0.9.9[-dev] for
     non-experimental use of the ECC ciphersuites to get TLS extension
     support, which is required for curve and point format negotiation
     to avoid potential handshake problems.
     [Bodo Moeller]

  *) Disable rogue ciphersuites:

      - SSLv2 0x08 0x00 0x80 ("RC4-64-MD5")
      - SSLv3/TLSv1 0x00 0x61 ("EXP1024-RC2-CBC-MD5")
      - SSLv3/TLSv1 0x00 0x60 ("EXP1024-RC4-MD5")

     The latter two were purportedly from
     draft-ietf-tls-56-bit-ciphersuites-0[01].txt, but do not really
     appear there.

     Also deactivate the remaining ciphersuites from
     draft-ietf-tls-56-bit-ciphersuites-01.txt.  These are just as
     unofficial, and the ID has long expired.
     [Bodo Moeller]

  *) Fix RSA blinding Heisenbug (problems sometimes occured on
     dual-core machines) and other potential thread-safety issues.
     [Bodo Moeller]

  *) Add the symmetric cipher Camellia (128-bit, 192-bit, 256-bit key
     versions), which is now available for royalty-free use
     (see http://info.isl.ntt.co.jp/crypt/eng/info/chiteki.html).
     Also, add Camellia TLS ciphersuites from RFC 4132.

     To minimize changes between patchlevels in the OpenSSL 0.9.8
     series, Camellia remains excluded from compilation unless OpenSSL
     is configured with 'enable-camellia'.
     [NTT]

  *) Disable the padding bug check when compression is in use. The padding
     bug check assumes the first packet is of even length, this is not
     necessarily true if compresssion is enabled and can result in false
     positives causing handshake failure. The actual bug test is ancient
     code so it is hoped that implementations will either have fixed it by
     now or any which still have the bug do not support compression.
     [Steve Henson]

 Changes between 0.9.8a and 0.9.8b  [04 May 2006]

  *) When applying a cipher rule check to see if string match is an explicit
     cipher suite and only match that one cipher suite if it is.
     [Steve Henson]

  *) Link in manifests for VC++ if needed.
     [Austin Ziegler <halostatue@gmail.com>]

  *) Update support for ECC-based TLS ciphersuites according to
     draft-ietf-tls-ecc-12.txt with proposed changes (but without
     TLS extensions, which are supported starting with the 0.9.9
     branch, not in the OpenSSL 0.9.8 branch).
     [Douglas Stebila]

  *) New functions EVP_CIPHER_CTX_new() and EVP_CIPHER_CTX_free() to support
     opaque EVP_CIPHER_CTX handling.
     [Steve Henson]

  *) Fixes and enhancements to zlib compression code. We now only use
     "zlib1.dll" and use the default __cdecl calling convention on Win32
     to conform with the standards mentioned here:
           http://www.zlib.net/DLL_FAQ.txt
     Static zlib linking now works on Windows and the new --with-zlib-include
     --with-zlib-lib options to Configure can be used to supply the location
     of the headers and library. Gracefully handle case where zlib library
     can't be loaded.
     [Steve Henson]

  *) Several fixes and enhancements to the OID generation code. The old code
     sometimes allowed invalid OIDs (1.X for X >= 40 for example), couldn't
     handle numbers larger than ULONG_MAX, truncated printing and had a
     non standard OBJ_obj2txt() behaviour.
     [Steve Henson]

  *) Add support for building of engines under engine/ as shared libraries
     under VC++ build system.
     [Steve Henson]

  *) Corrected the numerous bugs in the Win32 path splitter in DSO.
     Hopefully, we will not see any false combination of paths any more.
     [Richard Levitte]

 Changes between 0.9.8 and 0.9.8a  [11 Oct 2005]

  *) Remove the functionality of SSL_OP_MSIE_SSLV2_RSA_PADDING
     (part of SSL_OP_ALL).  This option used to disable the
     countermeasure against man-in-the-middle protocol-version
     rollback in the SSL 2.0 server implementation, which is a bad
     idea.  (CVE-2005-2969)

     [Bodo Moeller; problem pointed out by Yutaka Oiwa (Research Center
     for Information Security, National Institute of Advanced Industrial
     Science and Technology [AIST], Japan)]

  *) Add two function to clear and return the verify parameter flags.
     [Steve Henson]

  *) Keep cipherlists sorted in the source instead of sorting them at
     runtime, thus removing the need for a lock.
     [Nils Larsch]

  *) Avoid some small subgroup attacks in Diffie-Hellman.
     [Nick Mathewson and Ben Laurie]

  *) Add functions for well-known primes.
     [Nick Mathewson]

  *) Extended Windows CE support.
     [Satoshi Nakamura and Andy Polyakov]

  *) Initialize SSL_METHOD structures at compile time instead of during
     runtime, thus removing the need for a lock.
     [Steve Henson]

  *) Make PKCS7_decrypt() work even if no certificate is supplied by
     attempting to decrypt each encrypted key in turn. Add support to
     smime utility.
     [Steve Henson]

 Changes between 0.9.7h and 0.9.8  [05 Jul 2005]

  [NB: OpenSSL 0.9.7i and later 0.9.7 patch levels were released after
  OpenSSL 0.9.8.]

  *) Add libcrypto.pc and libssl.pc for those who feel they need them.
     [Richard Levitte]

  *) Change CA.sh and CA.pl so they don't bundle the CSR and the private
     key into the same file any more.
     [Richard Levitte]

  *) Add initial support for Win64, both IA64 and AMD64/x64 flavors.
     [Andy Polyakov]

  *) Add -utf8 command line and config file option to 'ca'.
     [Stefan <stf@udoma.org]

  *) Removed the macro des_crypt(), as it seems to conflict with some
     libraries.  Use DES_crypt().
     [Richard Levitte]

  *) Correct naming of the 'chil' and '4758cca' ENGINEs. This
     involves renaming the source and generated shared-libs for
     both. The engines will accept the corrected or legacy ids
     ('ncipher' and '4758_cca' respectively) when binding. NB,
     this only applies when building 'shared'.
     [Corinna Vinschen <vinschen@redhat.com> and Geoff Thorpe]

  *) Add attribute functions to EVP_PKEY structure. Modify
     PKCS12_create() to recognize a CSP name attribute and
     use it. Make -CSP option work again in pkcs12 utility.
     [Steve Henson]

  *) Add new functionality to the bn blinding code:
     - automatic re-creation of the BN_BLINDING parameters after
       a fixed number of uses (currently 32)
     - add new function for parameter creation
     - introduce flags to control the update behaviour of the
       BN_BLINDING parameters
     - hide BN_BLINDING structure
     Add a second BN_BLINDING slot to the RSA structure to improve
     performance when a single RSA object is shared among several
     threads.
     [Nils Larsch]

  *) Add support for DTLS.
     [Nagendra Modadugu <nagendra@cs.stanford.edu> and Ben Laurie]

  *) Add support for DER encoded private keys (SSL_FILETYPE_ASN1)
     to SSL_CTX_use_PrivateKey_file() and SSL_use_PrivateKey_file()
     [Walter Goulet]

  *) Remove buggy and incompletet DH cert support from
     ssl/ssl_rsa.c and ssl/s3_both.c
     [Nils Larsch]

  *) Use SHA-1 instead of MD5 as the default digest algorithm for
     the apps/openssl applications.
     [Nils Larsch]

  *) Compile clean with "-Wall -Wmissing-prototypes
     -Wstrict-prototypes -Wmissing-declarations -Werror". Currently
     DEBUG_SAFESTACK must also be set.
     [Ben Laurie]

  *) Change ./Configure so that certain algorithms can be disabled by default.
     The new counterpiece to "no-xxx" is "enable-xxx".

     The patented RC5 and MDC2 algorithms will now be disabled unless
     "enable-rc5" and "enable-mdc2", respectively, are specified.

     (IDEA remains enabled despite being patented.  This is because IDEA
     is frequently required for interoperability, and there is no license
     fee for non-commercial use.  As before, "no-idea" can be used to
     avoid this algorithm.)

     [Bodo Moeller]

  *) Add processing of proxy certificates (see RFC 3820).  This work was
     sponsored by KTH (The Royal Institute of Technology in Stockholm) and
     EGEE (Enabling Grids for E-science in Europe).
     [Richard Levitte]

  *) RC4 performance overhaul on modern architectures/implementations, such
     as Intel P4, IA-64 and AMD64.
     [Andy Polyakov]

  *) New utility extract-section.pl. This can be used specify an alternative
     section number in a pod file instead of having to treat each file as
     a separate case in Makefile. This can be done by adding two lines to the
     pod file:

     =for comment openssl_section:XXX

     The blank line is mandatory.

     [Steve Henson]

  *) New arguments -certform, -keyform and -pass for s_client and s_server
     to allow alternative format key and certificate files and passphrase
     sources.
     [Steve Henson]

  *) New structure X509_VERIFY_PARAM which combines current verify parameters,
     update associated structures and add various utility functions.

     Add new policy related verify parameters, include policy checking in 
     standard verify code. Enhance 'smime' application with extra parameters
     to support policy checking and print out.
     [Steve Henson]

  *) Add a new engine to support VIA PadLock ACE extensions in the VIA C3
     Nehemiah processors. These extensions support AES encryption in hardware
     as well as RNG (though RNG support is currently disabled).
     [Michal Ludvig <michal@logix.cz>, with help from Andy Polyakov]

  *) Deprecate BN_[get|set]_params() functions (they were ignored internally).
     [Geoff Thorpe]

  *) New FIPS 180-2 algorithms, SHA-224/-256/-384/-512 are implemented.
     [Andy Polyakov and a number of other people]

  *) Improved PowerPC platform support. Most notably BIGNUM assembler
     implementation contributed by IBM.
     [Suresh Chari, Peter Waltenberg, Andy Polyakov]

  *) The new 'RSA_generate_key_ex' function now takes a BIGNUM for the public
     exponent rather than 'unsigned long'. There is a corresponding change to
     the new 'rsa_keygen' element of the RSA_METHOD structure.
     [Jelte Jansen, Geoff Thorpe]

  *) Functionality for creating the initial serial number file is now
     moved from CA.pl to the 'ca' utility with a new option -create_serial.

     (Before OpenSSL 0.9.7e, CA.pl used to initialize the serial
     number file to 1, which is bound to cause problems.  To avoid
     the problems while respecting compatibility between different 0.9.7
     patchlevels, 0.9.7e  employed 'openssl x509 -next_serial' in
     CA.pl for serial number initialization.  With the new release 0.9.8,
     we can fix the problem directly in the 'ca' utility.)
     [Steve Henson]

  *) Reduced header interdepencies by declaring more opaque objects in
     ossl_typ.h. As a consequence, including some headers (eg. engine.h) will
     give fewer recursive includes, which could break lazy source code - so
     this change is covered by the OPENSSL_NO_DEPRECATED symbol. As always,
     developers should define this symbol when building and using openssl to
     ensure they track the recommended behaviour, interfaces, [etc], but
     backwards-compatible behaviour prevails when this isn't defined.
     [Geoff Thorpe]

  *) New function X509_POLICY_NODE_print() which prints out policy nodes.
     [Steve Henson]

  *) Add new EVP function EVP_CIPHER_CTX_rand_key and associated functionality.
     This will generate a random key of the appropriate length based on the 
     cipher context. The EVP_CIPHER can provide its own random key generation
     routine to support keys of a specific form. This is used in the des and 
     3des routines to generate a key of the correct parity. Update S/MIME
     code to use new functions and hence generate correct parity DES keys.
     Add EVP_CHECK_DES_KEY #define to return an error if the key is not 
     valid (weak or incorrect parity).
     [Steve Henson]

  *) Add a local set of CRLs that can be used by X509_verify_cert() as well
     as looking them up. This is useful when the verified structure may contain
     CRLs, for example PKCS#7 signedData. Modify PKCS7_verify() to use any CRLs
     present unless the new PKCS7_NO_CRL flag is asserted.
     [Steve Henson]

  *) Extend ASN1 oid configuration module. It now additionally accepts the
     syntax:

     shortName = some long name, 1.2.3.4
     [Steve Henson]

  *) Reimplemented the BN_CTX implementation. There is now no more static
     limitation on the number of variables it can handle nor the depth of the
     "stack" handling for BN_CTX_start()/BN_CTX_end() pairs. The stack
     information can now expand as required, and rather than having a single
     static array of bignums, BN_CTX now uses a linked-list of such arrays
     allowing it to expand on demand whilst maintaining the usefulness of
     BN_CTX's "bundling".
     [Geoff Thorpe]

  *) Add a missing BN_CTX parameter to the 'rsa_mod_exp' callback in RSA_METHOD
     to allow all RSA operations to function using a single BN_CTX.
     [Geoff Thorpe]

  *) Preliminary support for certificate policy evaluation and checking. This
     is initially intended to pass the tests outlined in "Conformance Testing
     of Relying Party Client Certificate Path Processing Logic" v1.07.
     [Steve Henson]

  *) bn_dup_expand() has been deprecated, it was introduced in 0.9.7 and
     remained unused and not that useful. A variety of other little bignum
     tweaks and fixes have also been made continuing on from the audit (see
     below).
     [Geoff Thorpe]

  *) Constify all or almost all d2i, c2i, s2i and r2i functions, along with
     associated ASN1, EVP and SSL functions and old ASN1 macros.
     [Richard Levitte]

  *) BN_zero() only needs to set 'top' and 'neg' to zero for correct results,
     and this should never fail. So the return value from the use of
     BN_set_word() (which can fail due to needless expansion) is now deprecated;
     if OPENSSL_NO_DEPRECATED is defined, BN_zero() is a void macro.
     [Geoff Thorpe]

  *) BN_CTX_get() should return zero-valued bignums, providing the same
     initialised value as BN_new().