oweals/openssl.git
8 years agoFix heartbeat_test
Matt Caswell [Wed, 2 Nov 2016 15:36:06 +0000 (15:36 +0000)]
Fix heartbeat_test

The heartbeat_test reaches into the internals of libssl and calls some
internal functions. It then checks the return value to check its what it
expected. However commit fa4c37457 changed the return value of these
internal functions, and now the test is failing.

The solution is to update the test to look for the new return value.

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoSecure our notification email.
Richard Levitte [Tue, 1 Nov 2016 23:09:03 +0000 (00:09 +0100)]
Secure our notification email.

Forks will have to define their own

Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1821)
(cherry picked from commit 5e28b1c1e048eef600dc49820934a5e1531186d4)

8 years agoFix grammar-o in CONTRIBUTING
Benjamin Kaduk [Mon, 26 Sep 2016 20:30:42 +0000 (15:30 -0500)]
Fix grammar-o in CONTRIBUTING

Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1625)
(cherry picked from commit e4d94269a5a41594852dc60716500580f1d47cef)

8 years agoAdd $(EX_LIBS) to the LIBDEPS for libgost.so, just as for all other engines
Richard Levitte [Fri, 28 Oct 2016 21:57:58 +0000 (23:57 +0200)]
Add $(EX_LIBS) to the LIBDEPS for libgost.so, just as for all other engines

Without this, any linker flag the user gave when configuring are ignored.

Reviewed-by: Andy Polyakov <appro@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1803)

8 years agoImplement length checks as a macro
Matt Caswell [Tue, 25 Oct 2016 10:10:56 +0000 (11:10 +0100)]
Implement length checks as a macro

Replace the various length checks in the extension code with a macro to
simplify the logic.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoEnsure we have length checks for all extensions
Matt Caswell [Fri, 14 Oct 2016 12:07:00 +0000 (13:07 +0100)]
Ensure we have length checks for all extensions

The previous commit inspired a review of all the length checks for the
extension adding code. This adds more robust checks and adds checks where
some were missing previously. The real solution for this is to use WPACKET
which is currently in master - but that cannot be applied to release
branches.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoFix length check writing status request extension
Matt Caswell [Fri, 14 Oct 2016 10:49:06 +0000 (11:49 +0100)]
Fix length check writing status request extension

The status request extension did not correctly check its length, meaning
that writing the extension could go 2 bytes beyond the buffer size. In
practice this makes little difference because, due to logic in buffer.c the
buffer is actually over allocated by approximately 5k!

Issue reported by Guido Vranken.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoFix a double free in ca command line
Matt Caswell [Mon, 10 Oct 2016 15:53:11 +0000 (16:53 +0100)]
Fix a double free in ca command line

Providing a spkac file with no default section causes a double free.

Thanks to Brian Carpenter for reporting this issue.

Reviewed-by: Kurt Roeckx <kurt@openssl.org>
(cherry picked from commit 229bd12487f8576fc088dc4f641950ac33c62033)

8 years agoA zero return from BIO_read/BIO_write() could be retryable
Matt Caswell [Fri, 21 Oct 2016 13:49:33 +0000 (14:49 +0100)]
A zero return from BIO_read/BIO_write() could be retryable

A zero return from BIO_read()/BIO_write() could mean that an IO operation
is retryable. A zero return from SSL_read()/SSL_write() means that the
connection has been closed down (either cleanly or not). Therefore we
should not propagate a zero return value from BIO_read()/BIO_write() back
up the stack to SSL_read()/SSL_write(). This could result in a retryable
failure being treated as fatal.

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoFix typo (reported by Matthias St. Pierre)
Rich Salz [Wed, 26 Oct 2016 15:48:43 +0000 (11:48 -0400)]
Fix typo (reported by Matthias St. Pierre)

Reviewed-by: Tim Hudson <tjh@openssl.org>
(cherry picked from commit 12a7715e3daed439e46cbed461d2a3d9dfd37c0f)

8 years agoFix leak of secrecy in ecdh_compute_key()
Dr. Matthias St. Pierre [Sat, 15 Oct 2016 22:53:33 +0000 (00:53 +0200)]
Fix leak of secrecy in ecdh_compute_key()

A temporary buffer containing g^xy was not cleared in ecdh_compute_key()
before freeing it, so the shared secret was leaked in memory.

Reviewed-by: Kurt Roeckx <kurt@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoCorrectly find all critical CRL extensions
Rich Salz [Sat, 22 Oct 2016 07:53:47 +0000 (03:53 -0400)]
Correctly find all critical CRL extensions

Unhandled critical CRL extensions were not detected if they appeared
after the handled ones.  (GitHub issue 1757).  Thanks to John Chuah
for reporting this.

Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1769)

8 years agoremove redundant zero assignments
Cristian Stoica [Wed, 17 Aug 2016 11:55:57 +0000 (14:55 +0300)]
remove redundant zero assignments

The structure has already been initialized to zero with memset.
See also commit 64b25758edca688a30f02c260262150f7ad0bc7d (remove 0 assignments)

Signed-off-by: Cristian Stoica <cristian.stoica@nxp.com>
CLA: trivial
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1463)

8 years ago%p takes void*, so make sure to cast arguments to void*
Richard Levitte [Thu, 20 Oct 2016 07:07:06 +0000 (09:07 +0200)]
%p takes void*, so make sure to cast arguments to void*

This avoids failures when configuring with --strict-warnings

Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1749)

8 years agoapps: make setup_engine() and release_engine() available always
Richard Levitte [Wed, 19 Oct 2016 17:46:38 +0000 (19:46 +0200)]
apps: make setup_engine() and release_engine() available always

This removes some #ifndef clutter.

Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1644)

8 years agoIf an engine comes up explicitely, it must also come down explicitely
Richard Levitte [Wed, 28 Sep 2016 22:40:20 +0000 (00:40 +0200)]
If an engine comes up explicitely, it must also come down explicitely

In apps/apps.c, one can set up an engine with setup_engine().
However, we freed the structural reference immediately, which means
that for engines that don't already have a structural reference
somewhere else (because it has registered at least one cipher or digest
algorithm method, and therefore gets a functional reference through the
ENGINE_set_default() call), we end up returning an invalid reference.

Instead, the function release_engine() is added, and called at the end
of the routines that call setup_engine().

Originally, the ENGINE API wasn't designed for this to happen, an
engine had to register at least one algorithm method, and was
especially expected to register the algorithms corresponding to the
key types that could be stored and hidden in hardware.  However, it
turns out that some engines will not register those algorithms with
the ENGINE_set_{algo}, ENGINE_set_cipher or ENGINE_set_digest
functions, as they only want the methods to be used for keys, not as
general crypto accelerator methods.  That may cause ENGINE_set_default()
to do nothing, and no functional reference is therefore made, leading
to a premature deallocation of the engine and it thereby becoming
unavailable when trying to fetch a key.

Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1644)

8 years agoFix no-des
Richard Levitte [Wed, 19 Oct 2016 20:54:06 +0000 (22:54 +0200)]
Fix no-des

Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1748)

8 years agoMake 'openssl prime ""' not segfault
Richard Levitte [Thu, 6 Oct 2016 07:31:34 +0000 (09:31 +0200)]
Make 'openssl prime ""' not segfault

Fixes RT#4699

Reviewed-by: Tim Hudson <tjh@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1668)

8 years agoFix strict-warnings build
Patrick Steuer [Mon, 17 Oct 2016 08:30:33 +0000 (10:30 +0200)]
Fix strict-warnings build

crypto/evp/e_aes.c: Types of inp and out parameters of AES_xts_en/decrypt
functions need to be changed from char to unsigned char to avoid build
error due to -Werror=incompatible-pointer-types.

crypto/aes/asm/aes-s390x.pl: Comments need to reflect the above change.

Signed-off-by: Patrick Steuer <psteuer@mail.de>
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
CLA: trivial

8 years agoFix strict-warnings build
Patrick Steuer [Mon, 17 Oct 2016 08:24:49 +0000 (10:24 +0200)]
Fix strict-warnings build

crypto/s390xcap.c: cryptlib.h needs to be included for OPENSSL_cpuid_setup
function prototype is located there to avoid build error due to
-Werror=missing-prototypes.

Signed-off-by: Patrick Steuer <psteuer@mail.de>
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
CLA: trivial

8 years agoFix signatures of EVP_Digest{Sign,Verify}Update
Steven Fackler [Sat, 15 Oct 2016 20:01:25 +0000 (13:01 -0700)]
Fix signatures of EVP_Digest{Sign,Verify}Update

These are implemented as macros delegating to `EVP_DigestUpdate`, which
takes a `size_t` as its third argument, not an `unsigned int`.

CLA: trivial

Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit 8bdce8d160e29b4e1b80fec31f618d85d8c2b7a8)

8 years agoEnsure we handle len == 0 in ERR_err_string_n
Matt Caswell [Wed, 12 Oct 2016 15:43:03 +0000 (16:43 +0100)]
Ensure we handle len == 0 in ERR_err_string_n

If len == 0 in a call to ERR_error_string_n() then we can read beyond the
end of the buffer. Really applications should not be calling this function
with len == 0, but we shouldn't be letting it through either!

Thanks to Agostino Sarubbo for reporting this issue. Agostino's blog on
this issue is available here:
https://blogs.gentoo.org/ago/2016/10/14/openssl-libcrypto-stack-based-buffer-overflow-in-err_error_string_n-err-c/

Reviewed-by: Richard Levitte <levitte@openssl.org>
(cherry picked from commit e5c1361580d8de79682958b04a5f0d262e680f8b)

8 years agoDegrade 3DES to MEDIUM in SSL2
Vitezslav Cizek [Mon, 10 Oct 2016 14:41:57 +0000 (16:41 +0200)]
Degrade 3DES to MEDIUM in SSL2

The SWEET32 fix moved 3DES from HIGH to MEDIUM, but omitted SSL2.

CLA: trivial
Reviewed-by: Kurt Roeckx <kurt@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1683)

8 years agoRT is put out to pasture
Rich Salz [Wed, 12 Oct 2016 19:49:06 +0000 (15:49 -0400)]
RT is put out to pasture

Reviewed-by: Tim Hudson <tjh@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1702)
(cherry picked from commit 7954dced19a7e59e7055eab95a981fa943c7d100)

8 years agoAdd missing error string for SSL_R_TOO_MANY_WARN_ALERTS
Kurt Cancemi [Thu, 22 Sep 2016 22:05:37 +0000 (18:05 -0400)]
Add missing error string for SSL_R_TOO_MANY_WARN_ALERTS

Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoapps/apps.c: initialize and de-initialize engine around key loading
Richard Levitte [Wed, 28 Sep 2016 19:28:00 +0000 (21:28 +0200)]
apps/apps.c: initialize and de-initialize engine around key loading

Before loading a key from an engine, it may need to be initialized.
When done loading the key, we must de-initialize the engine.
(if the engine is already initialized somehow, only the reference
counter will be incremented then decremented)

Reviewed-by: Stephen Henson <steve@openssl.org>
(cherry picked from commit 49e476a5382602d0bad1139d6f1f66ddbc7959d6)

8 years agoRevert "Call ENGINE_init() before trying to use keys from engine"
Rich Salz [Wed, 28 Sep 2016 18:39:32 +0000 (14:39 -0400)]
Revert "Call ENGINE_init() before trying to use keys from engine"

This reverts commit 4badd2b3c29c2c6c551c737c07a429a53d9d1a0d.
This fails to call ENGINE_finish; an alternate fix is coming.

Reviewed-by: Dr. Stephen Henson <steve@openssl.org>
8 years agoCall ENGINE_init() before trying to use keys from engine
David Woodhouse [Wed, 28 Sep 2016 13:31:22 +0000 (14:31 +0100)]
Call ENGINE_init() before trying to use keys from engine

Things like 'openssl s_client' only ever worked with keys from an engine
which provided a default generic method for some key type — because it
called ENGINE_set_default() and that ended up being an implicit
initialisation and functional refcount.

But an engine which doesn't provide generic methods doesn't get
initialised, and then when you try to use it you get an error:

cannot load client certificate private key file from engine
140688147056384:error:26096075:engine routines:ENGINE_load_private_key:not initialised:crypto/engine/eng_pkey.c:66:
unable to load client certificate private key file

cf. https://github.com/OpenSC/libp11/issues/107 (in which we discover
that engine_pkcs11 *used* to provide generic methods that OpenSSL would
try to use for ephemeral DH keys when negotiating ECDHE cipher suites in
TLS, and that didn't work out very well.)

Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1640)

8 years agoFix NEWS error
Matt Caswell [Mon, 26 Sep 2016 10:20:11 +0000 (11:20 +0100)]
Fix NEWS error

The NEWS file referenced the wrong CVE for 1.0.2

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoPrepare for 1.0.2k-dev
Matt Caswell [Mon, 26 Sep 2016 09:50:48 +0000 (10:50 +0100)]
Prepare for 1.0.2k-dev

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoPrepare for 1.0.2j release OpenSSL_1_0_2j
Matt Caswell [Mon, 26 Sep 2016 09:49:49 +0000 (10:49 +0100)]
Prepare for 1.0.2j release

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoUpdate CHANGES and NEWS for the new release
Matt Caswell [Mon, 26 Sep 2016 08:51:30 +0000 (09:51 +0100)]
Update CHANGES and NEWS for the new release

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoAdd some sanity checks when checking CRL scores
Matt Caswell [Mon, 22 Aug 2016 23:01:57 +0000 (00:01 +0100)]
Add some sanity checks when checking CRL scores

Note: this was accidentally omitted from OpenSSL 1.0.2 branch.
Without this fix any attempt to use CRLs will crash.

CVE-2016-7052

Thanks to Bruce Stephens and Thomas Jakobi for reporting this issue.

Reviewed-by: Stephen Henson <steve@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoFix build with no-nextprotoneg
Dirk Feytons [Thu, 22 Sep 2016 14:17:45 +0000 (16:17 +0200)]
Fix build with no-nextprotoneg

Add a missing ifdef. Same change is already present in master.

Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1100)

8 years agoFix typo introduced by a03f81f4
Rich Salz [Thu, 22 Sep 2016 12:47:45 +0000 (08:47 -0400)]
Fix typo introduced by a03f81f4

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoPrepare for 1.0.2j-dev
Matt Caswell [Thu, 22 Sep 2016 10:25:49 +0000 (11:25 +0100)]
Prepare for 1.0.2j-dev

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoPrepare for 1.0.2i release OpenSSL_1_0_2i
Matt Caswell [Thu, 22 Sep 2016 10:24:53 +0000 (11:24 +0100)]
Prepare for 1.0.2i release

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoUpdates CHANGES and NEWS for new release
Matt Caswell [Wed, 21 Sep 2016 20:59:49 +0000 (21:59 +0100)]
Updates CHANGES and NEWS for new release

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoAvoid KCI attack for GOST
Dmitry Belyavsky [Mon, 19 Sep 2016 15:05:53 +0000 (16:05 +0100)]
Avoid KCI attack for GOST

Russian GOST ciphersuites are vulnerable to the KCI attack because they use
long-term keys to establish the connection when ssl client authorization is
on. This change brings the GOST implementation into line with the latest
specs in order to avoid the attack. It should not break backwards
compatibility.

Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoFix a mem leak in NPN handling
Matt Caswell [Fri, 9 Sep 2016 09:53:39 +0000 (10:53 +0100)]
Fix a mem leak in NPN handling

If a server sent multiple NPN extensions in a single ClientHello then a
mem leak can occur. This will only happen where the client has requested
NPN in the first place. It does not occur during renegotiation. Therefore
the maximum that could be leaked in a single connection with a malicious
server is 64k (the maximum size of the ServerHello extensions section). As
this is client side, only occurs if NPN has been requested and does not
occur during renegotiation this is unlikely to be exploitable.

Issue reported by Shi Lei.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoFix OCSP Status Request extension unbounded memory growth
Matt Caswell [Fri, 9 Sep 2016 09:08:45 +0000 (10:08 +0100)]
Fix OCSP Status Request extension unbounded memory growth

A malicious client can send an excessively large OCSP Status Request
extension. If that client continually requests renegotiation,
sending a large OCSP Status Request extension each time, then there will
be unbounded memory growth on the server. This will eventually lead to a
Denial Of Service attack through memory exhaustion. Servers with a
default configuration are vulnerable even if they do not support OCSP.
Builds using the "no-ocsp" build time option are not affected.

I have also checked other extensions to see if they suffer from a similar
problem but I could not find any other issues.

CVE-2016-6304

Issue reported by Shi Lei.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agomk1mf.pl: check for no-tls1 here as well
Richard Levitte [Thu, 22 Sep 2016 08:01:38 +0000 (10:01 +0200)]
mk1mf.pl: check for no-tls1 here as well

Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoDon't allow too many consecutive warning alerts
Matt Caswell [Wed, 21 Sep 2016 13:48:16 +0000 (14:48 +0100)]
Don't allow too many consecutive warning alerts

Certain warning alerts are ignored if they are received. This can mean that
no progress will be made if one peer continually sends those warning alerts.
Implement a count so that we abort the connection if we receive too many.

Issue reported by Shi Lei.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoMake message buffer slightly larger than message.
Dr. Stephen Henson [Wed, 21 Sep 2016 12:26:01 +0000 (13:26 +0100)]
Make message buffer slightly larger than message.

Grow TLS/DTLS 16 bytes more than strictly necessary as a precaution against
OOB reads. In most cases this will have no effect because the message buffer
will be large enough already.

Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoUse SSL3_HM_HEADER_LENGTH instead of 4.
Dr. Stephen Henson [Wed, 21 Sep 2016 11:54:13 +0000 (12:54 +0100)]
Use SSL3_HM_HEADER_LENGTH instead of 4.

Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoRemove unnecessary check.
Dr. Stephen Henson [Wed, 21 Sep 2016 11:57:01 +0000 (12:57 +0100)]
Remove unnecessary check.

The overflow check will never be triggered because the
the n2l3 result is always less than 2^24.

Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoDcoument -alpn flag
Rich Salz [Wed, 21 Sep 2016 14:59:15 +0000 (10:59 -0400)]
Dcoument -alpn flag

Reviewed-by: Richard Levitte <levitte@openssl.org>
(cherry picked from commit 776e15f9393a9e3083bec60a8da376ce2fe1e97e)

8 years agoGH1555: Don't bump size on realloc failure
Rich Salz [Fri, 9 Sep 2016 14:52:59 +0000 (10:52 -0400)]
GH1555: Don't bump size on realloc failure

Reviewed-by: Richard Levitte <levitte@openssl.org>
(cherry picked from commit 6fcace45bda108ad4d3f95261494dd479720d92c)

8 years agoapps/apps.c: include sys/socket.h to declare recv()
Richard Levitte [Tue, 20 Sep 2016 16:43:24 +0000 (18:43 +0200)]
apps/apps.c: include sys/socket.h to declare recv()

Reviewed-by: Tim Hudson <tjh@openssl.org>
(cherry picked from commit a19228b7f4fc6fcb49713455b3caedbc24fb0b01)

8 years agoFix small OOB reads.
Dr. Stephen Henson [Sat, 17 Sep 2016 11:36:58 +0000 (12:36 +0100)]
Fix small OOB reads.

In ssl3_get_client_certificate, ssl3_get_server_certificate and
ssl3_get_certificate_request check we have enough room
before reading a length.

Thanks to Shi Lei (Gear Team, Qihoo 360 Inc.) for reporting these bugs.

CVE-2016-6306

Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoFix a missing NULL check in dsa_builtin_paramgen
Matt Caswell [Wed, 14 Sep 2016 12:27:59 +0000 (13:27 +0100)]
Fix a missing NULL check in dsa_builtin_paramgen

We should check the last BN_CTX_get() call to ensure that it isn't NULL
before we try and use any of the allocated BIGNUMs.

Issue reported by Shi Lei.

Reviewed-by: Richard Levitte <levitte@openssl.org>
(cherry picked from commit 1ff7425d6130380bb00d3e64739633a4b21b11a3)

8 years agoRT4669: dgst can only sign/verify one file
Richard Levitte [Tue, 20 Sep 2016 19:41:58 +0000 (21:41 +0200)]
RT4669: dgst can only sign/verify one file

Check arg count and print an error message.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoinitialize the RSA struct to 0.
Marcus Meissner [Tue, 6 Sep 2016 09:01:21 +0000 (11:01 +0200)]
initialize the RSA struct to 0.

This helps with program code linked against static builds accessing a uninitialized ->engine pointer.

CLA: none; trivial

Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1540)

8 years agoupdate default dependencies
Dr. Stephen Henson [Thu, 15 Sep 2016 22:49:41 +0000 (23:49 +0100)]
update default dependencies

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoRevert "Abort on unrecognised warning alerts"
Matt Caswell [Thu, 15 Sep 2016 19:53:09 +0000 (20:53 +0100)]
Revert "Abort on unrecognised warning alerts"

This reverts commit 15d81749322c3498027105f8ee44e8c25479d475.

There were some unexpected side effects to this commit, e.g. in SSLv3 a
warning alert gets sent "no_certificate" if a client does not send a
Certificate during Client Auth. With the above commit this causes the
connection to abort, which is incorrect. There may be some other edge cases
like this so we need to have a rethink on this.

Reviewed-by: Tim Hudson <tjh@openssl.org>
8 years agoFinally, make sure vms_term_sock.c is built
Richard Levitte [Wed, 14 Sep 2016 21:37:53 +0000 (23:37 +0200)]
Finally, make sure vms_term_sock.c is built

Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoRefactor to avoid unnecessary preprocessor logic
Richard Levitte [Thu, 15 Sep 2016 07:45:57 +0000 (09:45 +0200)]
Refactor to avoid unnecessary preprocessor logic

Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoReformat to fit OpenSSL source code standards
Richard Levitte [Wed, 14 Sep 2016 18:54:30 +0000 (20:54 +0200)]
Reformat to fit OpenSSL source code standards

Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoRemove entirely unnecessary pointer size guards
Richard Levitte [Wed, 14 Sep 2016 18:53:06 +0000 (20:53 +0200)]
Remove entirely unnecessary pointer size guards

Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoAdd copyright and license on apps/vms_term_sock.[ch]
Richard Levitte [Wed, 14 Sep 2016 18:52:03 +0000 (20:52 +0200)]
Add copyright and license on apps/vms_term_sock.[ch]

Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoVSI submission: redirect terminal input through socket
Richard Levitte [Wed, 3 Aug 2016 19:16:43 +0000 (21:16 +0200)]
VSI submission: redirect terminal input through socket

This is needed, because on VMS, select() can only be used on sockets.  being
able to use select() on all kinds of file descriptors is unique to Unix.

So, the solution for VMS is to create a layer that translates input from
standard input to socket communication.

Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoFix memory leak on realloc error.
Dr. Stephen Henson [Wed, 14 Sep 2016 22:42:55 +0000 (23:42 +0100)]
Fix memory leak on realloc error.

Backport leak fix from master branch.

Thanks to Shi Lei (Gear Team, Qihoo 360 Inc.) for reporting this bug.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoFix memory leak on error.
Dr. Stephen Henson [Wed, 14 Sep 2016 22:54:12 +0000 (23:54 +0100)]
Fix memory leak on error.

Thanks to Shi Lei (Gear Team, Qihoo 360 Inc.) for reporting this bug.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoVMS: only use _realloc32 with /POINTER_SIZE=32
Richard Levitte [Tue, 6 Sep 2016 15:39:35 +0000 (17:39 +0200)]
VMS: only use _realloc32 with /POINTER_SIZE=32

This fixes the following error when building with no particular pointer size
is specified (implied 32 bit):

    static void *(*realloc_func) (void *, size_t) = realloc;
    ................................................^
    %CC-E-UNDECLARED, In the initializer for realloc_func, "_realloc32" is not declared.
    at line number 93 in file DEV:[OPENSSL102.crypto]mem.c;1

Reviewed-by: Andy Polyakov <appro@openssl.org>
8 years agoAdd some sanity checks around usage of t_fromb64()
Matt Caswell [Thu, 8 Sep 2016 13:32:27 +0000 (14:32 +0100)]
Add some sanity checks around usage of t_fromb64()

The internal SRP function t_fromb64() converts from base64 to binary. It
does not validate that the size of the destination is sufficiently large -
that is up to the callers. In some places there was such a check, but not
in others.

Add an argument to t_fromb64() to provide the size of the destination
buffer and validate that we don't write too much data. Also add some sanity
checks to the callers where appropriate.

With thanks to Shi Lei for reporting this issue.

Reviewed-by: Richard Levitte <levitte@openssl.org>
(cherry picked from commit 73f0df8331910d6726d45ecaab12bd93cc48b4e2)

8 years agoAbort on unrecognised warning alerts
Matt Caswell [Mon, 12 Sep 2016 10:04:51 +0000 (11:04 +0100)]
Abort on unrecognised warning alerts

A peer continually sending unrecognised warning alerts could mean that we
make no progress on a connection. We should abort rather than continuing if
we receive an unrecognised warning alert.

Thanks to Shi Lei for reporting this issue.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoAdd enginesdir to libcrypto.pc pkg-config file
Richard Levitte [Mon, 12 Sep 2016 14:29:39 +0000 (16:29 +0200)]
Add enginesdir to libcrypto.pc pkg-config file

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoFix memory leak on error.
Dr. Stephen Henson [Thu, 8 Sep 2016 14:10:32 +0000 (15:10 +0100)]
Fix memory leak on error.

Backport leak fix from master branch.

Thanks to Shi Lei for reporting this bug.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoMake update
Rich Salz [Fri, 9 Sep 2016 16:17:47 +0000 (12:17 -0400)]
Make update

Reviewed-by: Andy Polyakov <appro@openssl.org>
8 years agoIf errno is ENXIO in BSS_new_file(), set BIO_R_NO_SUCH_FILE
Richard Levitte [Thu, 8 Sep 2016 21:39:26 +0000 (23:39 +0200)]
If errno is ENXIO in BSS_new_file(), set BIO_R_NO_SUCH_FILE

VMS sets that errno when the device part of a file spec is malformed
or a logical name that doesn't exist.

Reviewed-by: Rich Salz <rsalz@openssl.org>
(cherry picked from commit e82e2186e93e9a678dd8c0c5ba084d21d27d4d62)

8 years agoAvoid EVP_PKEY_cmp() crash on EC keys without public component
David Woodhouse [Wed, 7 Sep 2016 15:53:18 +0000 (16:53 +0100)]
Avoid EVP_PKEY_cmp() crash on EC keys without public component

Some hardware devices don't provide the public EC_POINT data. The only
way for X509_check_private_key() to validate that the key matches a
given certificate is to actually perform a sign operation and then
verify it using the public key in the certificate.

Maybe that can come later, as discussed in issue 1532. But for now let's
at least make it fail gracefully and not crash.

GH: 1532

Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/1547)
(cherry picked from commit 92ed7fa575a80955f3bb6efefca9bf576a953586)

8 years agoMisc BN fixes
Rich Salz [Mon, 5 Sep 2016 22:08:43 +0000 (18:08 -0400)]
Misc BN fixes

Never output -0; make "negative zero" an impossibility.
Do better checking on BN_rand top/bottom requirements and #bits.
Update doc.
Ignoring trailing garbage in BN_asc2bn.

Port this commit from boringSSL: https://boringssl.googlesource.com/boringssl/+/899b9b19a4cd3fe526aaf5047ab9234cdca19f7d%5E!/
        Ensure |BN_div| never gives negative zero in the no_branch code.

        Have |bn_correct_top| fix |bn->neg| if the input is zero so that we
        don't have negative zeros lying around.

        Thanks to Brian Smith for noticing.

Reviewed-by: Richard Levitte <levitte@openssl.org>
(cherry picked from commit 01c09f9fde5793e0b3712d602b02e2aed4908e8d)
(Some manual work required)

8 years agocrypto/bn/*: x86[_64] division instruction doesn't handle constants, change constrain...
Andy Polyakov [Sat, 27 Aug 2016 18:47:57 +0000 (20:47 +0200)]
crypto/bn/*: x86[_64] division instruction doesn't handle constants, change constraint from 'g' to 'r'.

Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Tim Hudson <tjh@openssl.org>
(cherry picked from commit 68b4a6e91f5acd42489bb9d1c580acc5ae457cad)

8 years agoEnsure the CertStatus message adds a DTLS message header where needed
Matt Caswell [Tue, 30 Aug 2016 14:06:01 +0000 (15:06 +0100)]
Ensure the CertStatus message adds a DTLS message header where needed

The function tls_construct_cert_status() is called by both TLS and DTLS
code. However it only ever constructed a TLS message header for the message
which obviously failed in DTLS.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoSRP_create_verifier does not check for NULL before OPENSSL_cleanse
Matt Caswell [Fri, 26 Aug 2016 14:14:24 +0000 (15:14 +0100)]
SRP_create_verifier does not check for NULL before OPENSSL_cleanse

OPENSSL_cleanse() does not validate its input parameter for NULL so
SRP_create_verifier() should do so instead. Otherwise a segfault will
result.

Alternative solution to GitHub PR#1006

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoImprove the definition of STITCHED_CALL in e_rc4_hmac_md5.c
Richard Levitte [Fri, 26 Aug 2016 07:59:55 +0000 (09:59 +0200)]
Improve the definition of STITCHED_CALL in e_rc4_hmac_md5.c

The definition of STITCHED_CALL relies on OPENSSL_NO_ASM.  However,
when a configuration simply lacks the assembler implementation for RC4
(which is where we have implemented the stitched call), OPENSSL_NO_ASM
isn't implemented.  Better, then, to rely on specific macros that
indicated that RC4 (and MD5) are implemented in assembler.

For this to work properly, we must also make sure Configure adds the
definition of RC4_ASM among the C flags.

(partly cherry picked from commit 216e8d91033d237880cff7da0d02d46d47bae41b)

Reviewed-by: Andy Polyakov <appro@openssl.org>
8 years agoFix a few leaks in X509_REQ_to_X509.
FdaSilvaYY [Mon, 4 Apr 2016 22:33:41 +0000 (00:33 +0200)]
Fix a few leaks in X509_REQ_to_X509.
Fix a possible leak on NETSCAPE_SPKI_verify failure.

Backport of 0517538d1a39bc
Backport of f6c006ea76304a

Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoAdd basic test for Cisco DTLS1_BAD_VER and record replay handling
David Woodhouse [Wed, 3 Aug 2016 17:25:07 +0000 (18:25 +0100)]
Add basic test for Cisco DTLS1_BAD_VER and record replay handling

(Modified for 1.0.2 by adding selected PACKET_xx() functions and PRF, and
subsequent cleanup from commit eb633d03fe2db3666840dee8d0a2dbe491672dfc)

Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit 40425899200a3dea9ec3684d3eb80bcf50c99baf)

8 years agoFix ubsan 'left shift of negative value -1' error in satsub64be()
David Woodhouse [Tue, 2 Aug 2016 21:54:46 +0000 (22:54 +0100)]
Fix ubsan 'left shift of negative value -1' error in satsub64be()

Baroque, almost uncommented code triggers behaviour which is undefined
by the C standard. You might quite reasonably not care that the code was
broken on ones-complement machines, but if we support a ubsan build then
we need to at least pretend to care.

It looks like the special-case code for 64-bit big-endian is going to
behave differently (and wrongly) on wrap-around, because it treats the
values as signed. That seems wrong, and allows replay and other attacks.
Surely you need to renegotiate and start a new epoch rather than
wrapping around to sequence number zero again?

Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit 2e94723c1b5d8ab974645e83de90b248265af3cd)

8 years agoFix SSL_export_keying_material() for DTLS1_BAD_VER
David Woodhouse [Fri, 8 Jul 2016 19:46:07 +0000 (20:46 +0100)]
Fix SSL_export_keying_material() for DTLS1_BAD_VER

Commit d8e8590e ("Fix missing return value checks in SCTP") made the
DTLS handshake fail, even for non-SCTP connections, if
SSL_export_keying_material() fails. Which it does, for DTLS1_BAD_VER.

Apply the trivial fix to make it succeed, since there's no real reason
why it shouldn't even though we never need it.

Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit c8a18468caef4d62778381be0acdadc8a88d6e51)

8 years agoFix the no-tls1 option
Matt Caswell [Fri, 26 Aug 2016 12:11:17 +0000 (13:11 +0100)]
Fix the no-tls1 option

This also fixes no-tls which is an alias for no-tls1 in 1.0.2 (it is not
possible to do no-tls1_1 or no-tls1_2 in 1.0.2).

Because it is not possible to disable TLS1.1 or TLS1.2 it no longer follows
that disabling TLS1.0 should force the disabling of tlsext.

Also a few missing ifdef guards.

GitHub Iusse#935

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoec/asm/ecp_nistz256-x86_64.pl: /cmovb/cmovc/ as nasm doesn't recognize cmovb.
Andy Polyakov [Wed, 24 Aug 2016 15:13:09 +0000 (17:13 +0200)]
ec/asm/ecp_nistz256-x86_64.pl: /cmovb/cmovc/ as nasm doesn't recognize cmovb.

Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit d3034d31e7c04b334dd245504dd4f56e513ca115)

8 years agoec/ecp_nistz256: harmonize is_infinity with ec_GFp_simple_is_at_infinity.
Andy Polyakov [Fri, 19 Aug 2016 21:16:04 +0000 (23:16 +0200)]
ec/ecp_nistz256: harmonize is_infinity with ec_GFp_simple_is_at_infinity.

RT#4625

Reviewed-by: Rich Salz <rsalz@openssl.org>
(cherry picked from commit e3057a57caf4274ea1fb074518e4714059dfcabf)

8 years agoec/asm/ecp_nistz256-x86_64.pl: addition to perform stricter reduction.
Andy Polyakov [Sat, 20 Aug 2016 20:04:21 +0000 (22:04 +0200)]
ec/asm/ecp_nistz256-x86_64.pl: addition to perform stricter reduction.

Addition was not preserving inputs' property of being fully reduced.

Thanks to Brian Smith for reporting this.

Reviewed-by: Rich Salz <rsalz@openssl.org>
(cherry picked from commit b62b2454fadfccaf5e055a1810d72174c2633b8f)

8 years agoAlways use session_ctx when removing a session
Todd Short [Thu, 26 May 2016 17:49:36 +0000 (13:49 -0400)]
Always use session_ctx when removing a session

Sessions are stored on the session_ctx, which doesn't change after
SSL_set_SSL_CTX().

Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoAvoid overflow in MDC2_Update()
Dr. Stephen Henson [Fri, 19 Aug 2016 22:28:29 +0000 (23:28 +0100)]
Avoid overflow in MDC2_Update()

Thanks to Shi Lei for reporting this issue.

CVE-2016-6303

Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit 55d83bf7c10c7b205fffa23fa7c3977491e56c07)

8 years agoSWEET32 (CVE-2016-2183): Move DES from HIGH to MEDIUM
Rich Salz [Thu, 18 Aug 2016 13:26:52 +0000 (09:26 -0400)]
SWEET32 (CVE-2016-2183): Move DES from HIGH to MEDIUM

Reviewed-by: Viktor Dukhovni <viktor@openssl.org>
Reviewed-by: Emilia Käsper <emilia@openssl.org>
8 years agoFix no-ec
Matt Caswell [Wed, 24 Aug 2016 08:23:14 +0000 (09:23 +0100)]
Fix no-ec

Use a ciphersuite in dtlstest that is not affected by no-* options.
Backport of commit fe34735c19.

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoSanity check ticket length.
Dr. Stephen Henson [Tue, 23 Aug 2016 17:14:54 +0000 (18:14 +0100)]
Sanity check ticket length.

If a ticket callback changes the HMAC digest to SHA512 the existing
sanity checks are not sufficient and an attacker could perform a DoS
attack with a malformed ticket. Add additional checks based on
HMAC size.

Thanks to Shi Lei for reporting this bug.

CVE-2016-6302

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agomk1mf: dtlstest needs ssltestlib, include it with a hack
Richard Levitte [Fri, 19 Aug 2016 14:53:54 +0000 (16:53 +0200)]
mk1mf: dtlstest needs ssltestlib, include it with a hack

We don't really have a mechanism to include other object files into a given
test program.  For now, a simple hack in mk1mf.pl will do.

RT#4653

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoDon't check for malloc failure twice.
David Benjamin [Sat, 20 Aug 2016 17:55:17 +0000 (13:55 -0400)]
Don't check for malloc failure twice.

a03f81f4ead24c234dc26e388d86a352685f3948 added a malloc failure check to
EVP_PKEY_keygen, but there already was one.

Signed-off-by: Kurt Roeckx <kurt@roeckx.be>
Reviewed-by: Rich Salz <rsalz@openssl.org>
GH: #1473

8 years agoFix overflow check in BN_bn2dec()
Kazuki Yamaguchi [Sun, 21 Aug 2016 17:36:36 +0000 (02:36 +0900)]
Fix overflow check in BN_bn2dec()

Fix an off by one error in the overflow check added by 07bed46f332fc
("Check for errors in BN_bn2dec()").

Reviewed-by: Stephen Henson <steve@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit 099e2968ed3c7d256cda048995626664082b1b30)

8 years agoRT2676: Reject RSA eponent if even or 1
Rich Salz [Mon, 22 Aug 2016 15:25:12 +0000 (11:25 -0400)]
RT2676: Reject RSA eponent if even or 1

Also, re-organize RSA check to use goto err.
Try all checks, not just stopping at first (via Richard Levitte)

Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(cherry picked from commit 464d59a5bb5811f7671e2bd37f41d610606b829d)

8 years agoVMS: Use strict refdef extern model when building library object files
Richard Levitte [Mon, 22 Aug 2016 08:17:27 +0000 (10:17 +0200)]
VMS: Use strict refdef extern model when building library object files

Most of the time, this isn't strictly needed.  However, in the default
extern model (called relaxed refdef), symbols are treated as weak
common objects unless they are initialised.  The librarian doesn't
include weak symbols in the (static) libraries, which renders them
invisible when linking a program with said those libraries, which is a
problem at times.

Using the strict refdef model is much more like standard C on all
other platforms, and thereby avoid the issues that come with the
relaxed refdef model.

Note: this doesn't apply to VAX C.  It's possible that this will make
OpenSSL building with VAX C difficult some time in the future if it
isn't already.  However, VAX C is a very old compiler that we don't
expect to see too often, as DEC C (a.k.a VMS C) should have replaced
it a long time ago.

Reviewed-by: Rich Salz <rsalz@openssl.org>
8 years agoGOST: rearrange code so it's more like C rather than C++
Richard Levitte [Mon, 22 Aug 2016 11:35:27 +0000 (13:35 +0200)]
GOST: rearrange code so it's more like C rather than C++

Some builds fail otherwise.

Reviewed-by: Matt Caswell <matt@openssl.org>
8 years agoMake 'openssl req -x509' more equivalent to 'openssl req -new'
Richard Levitte [Mon, 22 Aug 2016 13:22:17 +0000 (15:22 +0200)]
Make 'openssl req -x509' more equivalent to 'openssl req -new'

The following would fail, or rather, freeze:

    openssl genrsa -out rsa2048.pem 2048
    openssl req -x509 -key rsa2048.pem -keyform PEM -out cert.pem

In that case, the second command wants to read a certificate request
from stdin, because -x509 wasn't fully flagged as being for creating
something new.  This changes makes it fully flagged.

RT#4655

Reviewed-by: Andy Polyakov <appro@openssl.org>
8 years agobn/asm/x86[_64]-mont*.pl: implement slightly alternative page-walking.
Andy Polyakov [Wed, 16 Mar 2016 22:33:53 +0000 (23:33 +0100)]
bn/asm/x86[_64]-mont*.pl: implement slightly alternative page-walking.

Original strategy for page-walking was adjust stack pointer and then
touch pages in order. This kind of asks for double-fault, because
if touch fails, then signal will be delivered to frame above adjusted
stack pointer. But touching pages prior adjusting stack pointer would
upset valgrind. As compromise let's adjust stack pointer in pages,
touching top of the stack. This still asks for double-fault, but at
least prevents corruption of neighbour stack if allocation is to
overstep the guard page.

Also omit predict-non-taken hints as they reportedly trigger illegal
instructions in some VM setups.

Reviewed-by: Richard Levitte <levitte@openssl.org>
(cherry picked from commit 3ba1ef829cf3dd36eaa5e819258d90291c6a1027)

8 years agossltestlib: Tell compiler we don't care about the value when we don't
Richard Levitte [Mon, 22 Aug 2016 12:02:31 +0000 (14:02 +0200)]
ssltestlib: Tell compiler we don't care about the value when we don't

In mempacket_test_read(), we've already fetched the top value of the
stack, so when we shift the stack, we don't care for the value.  The
compiler needs to be told, or it will complain harshly when we tell it
to be picky.

Reviewed-by: Matt Caswell <matt@openssl.org>
(cherry picked from commit 1c288878af42650fbda911b702ae7b551a545b1c)

8 years agoPrevent DTLS Finished message injection
Matt Caswell [Thu, 30 Jun 2016 14:06:27 +0000 (15:06 +0100)]
Prevent DTLS Finished message injection

Follow on from CVE-2016-2179

The investigation and analysis of CVE-2016-2179 highlighted a related flaw.

This commit fixes a security "near miss" in the buffered message handling
code. Ultimately this is not currently believed to be exploitable due to
the reasons outlined below, and therefore there is no CVE for this on its
own.

The issue this commit fixes is a MITM attack where the attacker can inject
a Finished message into the handshake. In the description below it is
assumed that the attacker injects the Finished message for the server to
receive it. The attack could work equally well the other way around (i.e
where the client receives the injected Finished message).

The MITM requires the following capabilities:
- The ability to manipulate the MTU that the client selects such that it
is small enough for the client to fragment Finished messages.
- The ability to selectively drop and modify records sent from the client
- The ability to inject its own records and send them to the server

The MITM forces the client to select a small MTU such that the client
will fragment the Finished message. Ideally for the attacker the first
fragment will contain all but the last byte of the Finished message,
with the second fragment containing the final byte.

During the handshake and prior to the client sending the CCS the MITM
injects a plaintext Finished message fragment to the server containing
all but the final byte of the Finished message. The message sequence
number should be the one expected to be used for the real Finished message.

OpenSSL will recognise that the received fragment is for the future and
will buffer it for later use.

After the client sends the CCS it then sends its own Finished message in
two fragments. The MITM causes the first of these fragments to be
dropped. The OpenSSL server will then receive the second of the fragments
and reassemble the complete Finished message consisting of the MITM
fragment and the final byte from the real client.

The advantage to the attacker in injecting a Finished message is that
this provides the capability to modify other handshake messages (e.g.
the ClientHello) undetected. A difficulty for the attacker is knowing in
advance what impact any of those changes might have on the final byte of
the handshake hash that is going to be sent in the "real" Finished
message. In the worst case for the attacker this means that only 1 in
256 of such injection attempts will succeed.

It may be possible in some situations for the attacker to improve this such
that all attempts succeed. For example if the handshake includes client
authentication then the final message flight sent by the client will
include a Certificate. Certificates are ASN.1 objects where the signed
portion is DER encoded. The non-signed portion could be BER encoded and so
the attacker could re-encode the certificate such that the hash for the
whole handshake comes to a different value. The certificate re-encoding
would not be detectable because only the non-signed portion is changed. As
this is the final flight of messages sent from the client the attacker
knows what the complete hanshake hash value will be that the client will
send - and therefore knows what the final byte will be. Through a process
of trial and error the attacker can re-encode the certificate until the
modified handhshake also has a hash with the same final byte. This means
that when the Finished message is verified by the server it will be
correct in all cases.

In practice the MITM would need to be able to perform the same attack
against both the client and the server. If the attack is only performed
against the server (say) then the server will not detect the modified
handshake, but the client will and will abort the connection.
Fortunately, although OpenSSL is vulnerable to Finished message
injection, it is not vulnerable if *both* client and server are OpenSSL.
The reason is that OpenSSL has a hard "floor" for a minimum MTU size
that it will never go below. This minimum means that a Finished message
will never be sent in a fragmented form and therefore the MITM does not
have one of its pre-requisites. Therefore this could only be exploited
if using OpenSSL and some other DTLS peer that had its own and separate
Finished message injection flaw.

The fix is to ensure buffered messages are cleared on epoch change.

Reviewed-by: Richard Levitte <levitte@openssl.org>
8 years agoFix DTLS buffered message DoS attack
Matt Caswell [Thu, 30 Jun 2016 12:17:08 +0000 (13:17 +0100)]
Fix DTLS buffered message DoS attack

DTLS can handle out of order record delivery. Additionally since
handshake messages can be bigger than will fit into a single packet, the
messages can be fragmented across multiple records (as with normal TLS).
That means that the messages can arrive mixed up, and we have to
reassemble them. We keep a queue of buffered messages that are "from the
future", i.e. messages we're not ready to deal with yet but have arrived
early. The messages held there may not be full yet - they could be one
or more fragments that are still in the process of being reassembled.

The code assumes that we will eventually complete the reassembly and
when that occurs the complete message is removed from the queue at the
point that we need to use it.

However, DTLS is also tolerant of packet loss. To get around that DTLS
messages can be retransmitted. If we receive a full (non-fragmented)
message from the peer after previously having received a fragment of
that message, then we ignore the message in the queue and just use the
non-fragmented version. At that point the queued message will never get
removed.

Additionally the peer could send "future" messages that we never get to
in order to complete the handshake. Each message has a sequence number
(starting from 0). We will accept a message fragment for the current
message sequence number, or for any sequence up to 10 into the future.
However if the Finished message has a sequence number of 2, anything
greater than that in the queue is just left there.

So, in those two ways we can end up with "orphaned" data in the queue
that will never get removed - except when the connection is closed. At
that point all the queues are flushed.

An attacker could seek to exploit this by filling up the queues with
lots of large messages that are never going to be used in order to
attempt a DoS by memory exhaustion.

I will assume that we are only concerned with servers here. It does not
seem reasonable to be concerned about a memory exhaustion attack on a
client. They are unlikely to process enough connections for this to be
an issue.

A "long" handshake with many messages might be 5 messages long (in the
incoming direction), e.g. ClientHello, Certificate, ClientKeyExchange,
CertificateVerify, Finished. So this would be message sequence numbers 0
to 4. Additionally we can buffer up to 10 messages in the future.
Therefore the maximum number of messages that an attacker could send
that could get orphaned would typically be 15.

The maximum size that a DTLS message is allowed to be is defined by
max_cert_list, which by default is 100k. Therefore the maximum amount of
"orphaned" memory per connection is 1500k.

Message sequence numbers get reset after the Finished message, so
renegotiation will not extend the maximum number of messages that can be
orphaned per connection.

As noted above, the queues do get cleared when the connection is closed.
Therefore in order to mount an effective attack, an attacker would have
to open many simultaneous connections.

Issue reported by Quan Luo.

CVE-2016-2179

Reviewed-by: Richard Levitte <levitte@openssl.org>