Matt Caswell [Thu, 5 Feb 2015 13:59:16 +0000 (13:59 +0000)]
Add ticket length before buffering DTLS message
In ssl3_send_new_session_ticket the message to be sent is constructed. We
skip adding the length of the session ticket initially, then call
ssl_set_handshake_header, and finally go back and add in the length of the
ticket. Unfortunately, in DTLS, ssl_set_handshake_header also has the side
effect of buffering the message for subsequent retransmission if required.
By adding the ticket length after the call to ssl_set_handshake_header the
message that is buffered is incomplete, causing an invalid message to be
sent on retransmission.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Matt Caswell [Thu, 5 Feb 2015 13:54:37 +0000 (13:54 +0000)]
Ensure last_write_sequence is saved in DTLS1.2
In DTLS, immediately prior to epoch change, the write_sequence is supposed
to be stored in s->d1->last_write_sequence. The write_sequence is then reset
back to 00000000. In the event of retransmits of records from the previous
epoch, the last_write_sequence is restored. This commit fixes a bug in
DTLS1.2 where the write_sequence was being reset before last_write_sequence
was saved, and therefore retransmits are sent with incorrect sequence
numbers.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Richard Levitte [Tue, 24 Mar 2015 14:02:51 +0000 (15:02 +0100)]
Teach mkdef.pl to handle multiline declarations.
For the moment, this is specially crafted for DECLARE_DEPRECATED because
that's where we found the problem, but it can easily be expanded to other
types of special delarations when needed.
Disable loop checking when we retry verification with an alternative path.
This fixes the case where an intermediate CA is explicitly trusted and part
of the untrusted certificate list. By disabling loop checking for this case
the untrusted CA can be replaced by the explicitly trusted case and
verification will succeed.
If a set of certificates is supplied to OCSP_basic_verify use those in
addition to any present in the OCSP response as untrusted CAs when
verifying a certificate chain.
Move ASN.1 internals used across multiple directories into new internal
header file asn1_int.h remove crypto/Makefile hack which allowed other
directories to include "asn1_locl.h"
Richard Levitte [Tue, 24 Mar 2015 10:59:01 +0000 (11:59 +0100)]
Adjust include path
Thanks to a -I.., the path does work, at least on unix. However, this
doesn't work so well on VMS. Correcting the path to not rely on given
-I does work on both.
Remove old ASN.1 COMPAT type. This was meant as a temporary measure
so older ASN.1 code (from OpenSSL 0.9.6) still worked. It's a hack
which breaks constification and hopefully nothing uses it now, if
it ever did.
Matt Caswell [Thu, 5 Mar 2015 10:14:40 +0000 (10:14 +0000)]
Check libssl function returns
Mark most functions returning a result defined in any libssl header file
with __owur to warn if they are used without checking the return value.
Use -DUNUSED_RETURN compiler flag with gcc to activate these warnings.
Some functions returning a result are skipped if it is common and valid to
use these functions without checking the return value.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Matt Caswell [Fri, 6 Mar 2015 14:22:22 +0000 (14:22 +0000)]
Add -DDEBUG_UNUSED to --strict-warnings
In order to receive warnings on unused function return values the flag
-DDEBUG_UNUSED must be passed to the compiler. This change adds that for the
--strict-warnings Configure option.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Matt Caswell [Fri, 20 Mar 2015 15:10:16 +0000 (15:10 +0000)]
Don't check curves that haven't been sent
Don't check that the curve appears in the list of acceptable curves for the
peer, if they didn't send us such a list (RFC 4492 does not require that the
extension be sent).
Richard Godbee [Sat, 14 Mar 2015 04:23:21 +0000 (21:23 -0700)]
CRYPTO_128_unwrap(): Fix refactoring damage
crypto/modes/wrap128.c was heavily refactored to support AES Key Wrap
with Padding, and four bugs were introduced into CRYPTO_128_unwrap() at
that time:
- crypto_128_unwrap_raw()'s return value ('ret') is checked incorrectly,
and the function immediately returns 'ret' in (almost) all cases.
This makes the IV checking code later in the function unreachable, but
callers think the IV check succeeded since CRYPTO_128_unwrap()'s
return value is non-zero.
FIX: Return 0 (error) if crypto_128_unwrap_raw() returned 0 (error).
- crypto_128_unwrap_raw() writes the IV to the 'got_iv' buffer, not to
the first 8 bytes of the output buffer ('out') as the IV checking code
expects. This makes the IV check fail.
FIX: Compare 'iv' to 'got_iv', not 'out'.
- The data written to the output buffer ('out') is "cleansed" if the IV
check fails, but the code passes OPENSSL_cleanse() the input buffer
length ('inlen') instead of the number of bytes that
crypto_128_unwrap_raw() wrote to the output buffer ('ret'). This
means that OPENSSL_cleanse() could potentially write past the end of
'out'.
FIX: Change 'inlen' to 'ret' in the OPENSSL_cleanse() call.
- CRYPTO_128_unwrap() is returning the length of the input buffer
('inlen') instead of the number of bytes written to the output buffer
('ret'). This could cause the caller to read past the end of 'out'.
FIX: Return 'ret' instead of 'inlen' at the end of the function.
PR#3749
Reviewed-by: Stephen Henson <steve@openssl.org> Reviewed-by: Emilia Käsper <emilia@openssl.org>
Emilia Kasper [Fri, 27 Feb 2015 15:52:23 +0000 (16:52 +0100)]
PKCS#7: avoid NULL pointer dereferences with missing content
In PKCS#7, the ASN.1 content component is optional.
This typically applies to inner content (detached signatures),
however we must also handle unexpected missing outer content
correctly.
This patch only addresses functions reachable from parsing,
decryption and verification, and functions otherwise associated
with reading potentially untrusted data.
Correcting all low-level API calls requires further work.
CVE-2015-0289
Thanks to Michal Zalewski (Google) for reporting this issue.
Fix segmentation violation when ASN1_TYPE_cmp is passed a boolean type. This
can be triggered during certificate verification so could be a DoS attack
against a client or a server enabling client authentication.
CVE-2015-0286
Reviewed-by: Richard Levitte <levitte@openssl.org>
Matt Caswell [Tue, 10 Mar 2015 16:38:32 +0000 (16:38 +0000)]
Fix DHE Null CKE vulnerability
If client auth is used then a server can seg fault in the event of a DHE
cipher being used and a zero length ClientKeyExchange message being sent
by the client. This could be exploited in a DoS attack.
CVE-2015-1787
Reviewed-by: Richard Levitte <levitte@openssl.org>
Fix a bug where invalid PSS parameters are not rejected resulting in a
NULL pointer exception. This can be triggered during certificate
verification so could be a DoS attack against a client or a server
enabling client authentication.
Thanks to Brian Carpenter for reporting this issues.
Matt Caswell [Mon, 9 Mar 2015 16:09:04 +0000 (16:09 +0000)]
Fix Seg fault in DTLSv1_listen
The DTLSv1_listen function is intended to be stateless and processes
the initial ClientHello from many peers. It is common for user code to
loop over the call to DTLSv1_listen until a valid ClientHello is received
with an associated cookie. A defect in the implementation of DTLSv1_listen
means that state is preserved in the SSL object from one invokation to the
next that can lead to a segmentation fault. Erorrs processing the initial
ClientHello can trigger this scenario. An example of such an error could
be that a DTLS1.0 only client is attempting to connect to a DTLS1.2 only
server.
CVE-2015-0207
Reviewed-by: Richard Levitte <levitte@openssl.org>
Matt Caswell [Mon, 2 Mar 2015 09:27:10 +0000 (09:27 +0000)]
Multiblock corrupted pointer fix
OpenSSL 1.0.2 introduced the "multiblock" performance improvement. This
feature only applies on 64 bit x86 architecture platforms that support AES
NI instructions. A defect in the implementation of "multiblock" can cause
OpenSSL's internal write buffer to become incorrectly set to NULL when
using non-blocking IO. Typically, when the user application is using a
socket BIO for writing, this will only result in a failed connection.
However if some other BIO is used then it is likely that a segmentation
fault will be triggered, thus enabling a potential DoS attack.
CVE-2015-0290
Reviewed-by: Richard Levitte <levitte@openssl.org> Reviewed-by: Andy Polyakov <appro@openssl.org>
Richard Levitte [Tue, 17 Mar 2015 15:30:54 +0000 (16:30 +0100)]
Correct the request of debug builds
./config would translate -d into having the target get a 'debug-'
prefix, and then run './Configure LIST' to find out if such a
debugging target exists or not.
With the recent changes, the separate 'debug-foo' targets are
disappearing, and we're giving the normal targets debugging
capabilities instead. Unfortunately, './config' wasn't changed to
match this new behavior.
This change introduces the arguments '--debug' and '--release' - the
latter just for orthogonality - to ./Configure, and ./config now
treats -d by adding '--debug' to the options for ./Configure.
Matt Caswell [Fri, 13 Mar 2015 12:48:57 +0000 (12:48 +0000)]
Fix probable_prime over large shift
In the probable_prime() function we behave slightly different if the number
of bits we are interested in is <= BN_BITS2 (the num of bits in a BN_ULONG).
As part of the calculation we work out a size_limit as follows:
There is a problem though if bits == BN_BITS2. Shifting by that much causes
undefined behaviour. I did some tests. On my system BN_BITS2 == 64. So I
set bits to 64 and calculated the result of:
(((BN_ULONG)1) << bits)
I was expecting to get the result 0. I actually got 1! Strangely this...
(((BN_ULONG)0) << BN_BITS2)
...does equal 0! This means that, on my system at least, size_limit will be
off by 1 when bits == BN_BITS2.
This commit fixes the behaviour so that we always get consistent results.
Matt Caswell [Thu, 12 Mar 2015 15:59:07 +0000 (15:59 +0000)]
Fix unintended sign extension
The function CRYPTO_128_unwrap_pad uses an 8 byte AIV (Alternative Initial
Value). The least significant 4 bytes of this is placed into the local
variable |ptext_len|. This is done as follows:
aiv[4] is an unsigned char, but (aiv[4] << 24) is promoted to a *signed*
int - therefore we could end up shifting into the sign bit and end up with
a negative value. |ptext_len| is a size_t (typically 64-bits). If the
result of the shifts is negative then the upper bits of |ptext_len| will
all be 1.
This commit fixes the issue by explicitly casting to an unsigned int.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Matt Caswell [Thu, 12 Mar 2015 14:37:26 +0000 (14:37 +0000)]
Add sanity check to PRF
The function tls1_PRF counts the number of digests in use and partitions
security evenly between them. There always needs to be at least one digest
in use, otherwise this is an internal error. Add a sanity check for this.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Matt Caswell [Thu, 12 Mar 2015 12:54:44 +0000 (12:54 +0000)]
Fix memset call in stack.c
The function sk_zero is supposed to zero the elements held within a stack.
It uses memset to do this. However it calculates the size of each element
as being sizeof(char **) instead of sizeof(char *). This probably doesn't
make much practical difference in most cases, but isn't a portable
assumption.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Matt Caswell [Thu, 12 Mar 2015 11:25:03 +0000 (11:25 +0000)]
Move malloc fail checks closer to malloc
Move memory allocation failure checks closer to the site of the malloc in
dgst app. Only a problem if the debug flag is set...but still should be
fixed.
Richard Levitte [Mon, 16 Mar 2015 21:36:19 +0000 (22:36 +0100)]
Do not keep TABLE in version control.
TABLE was always a debugging tool, and permitted everyone to see the
effect of changes in the string-format configs. The hash-format
configs being much more readable, distributing TABLE becomes much less
necessary.
Being able to produce a TABLE is kept, however, as it still is a
useful debugging tool for configs, what with multi-level inheritance
and all.
Richard Levitte [Tue, 10 Mar 2015 21:04:44 +0000 (22:04 +0100)]
Rethink templates.
Because base templates express inheritance of values, the attribute is
renamed to 'inherit_from', and texts about this talk about 'inheritance(s)'
rather than base templates.
As they were previously implemented, base templates that were listed
together would override one another, the first one acting as defaults for
the next and so on.
However, it was pointed out that a strength of inheritance would be to
base configurations on several templates - for example one for CPU, one
for operating system and one for compiler - and that requires a different
way of combining those templates. With this change, inherited values
from several inheritances are concatenated by default (keep on reading).
Also, in-string templates with the double-curly syntax are removed,
replaced with the possibility to have a configuration value be a coderef
(i.e. a 'sub { /* your code goes here */ }') that gets the list of values
from all inheritances as the list @_. The result of executing such a
coderef on a list of values is assumed to become a string. ANY OTHER
FORM OF VALUE WILL CURRENTLY BREAK.
As a matter of fact, an attribute in the current config with no value is
assumed to have this coderef as value:
sub { join(' ', @_) }
While we're at it, rename debug-[cl]flags to debug_[cl]flags and
nodebug-[cl]flags to release_[cl]flags.
Richard Levitte [Fri, 6 Mar 2015 09:01:08 +0000 (10:01 +0100)]
Add base template processing.
Base templates are templates that are used to inherit from. They can
loosely be compared with parent class inheritance in object orientation.
They can be used for the same purpose as the variables with multi-field
strings are used in old-style string configurations.
Base templates are declared with the base_templates configuration
attribute, like so:
Note: The value of base_templates MUST be an array reference (an array
enclosed in square brackets).
Any configuration target can be used as a base template by another. It
is also possible to have a target that's a pure template and not meant to
be used directly as a configuration target. Such a target is marked with
the template configuration attribute, like so:
As part of this commit, all variables with multi-field strings have been
translated to pure templates. The variables currently remain since we
can't expect people to shift to hash table configurations immediately.
Richard Levitte [Fri, 6 Mar 2015 01:00:21 +0000 (02:00 +0100)]
Rewrite Configure to handle the target values as hash tables.
The reasoning is that configuration strings are hard to read and error
prone, and that a better way would be for them to be key => value hashes.
Configure is made to be able to handle target configuration values as a
string as well as a hash. It also does the best it can to combine a
"debug-foo" target with a "foo" target, given that they are similar
except for the cflags and lflags values. The latter are spliced into
options that are common for "debug-foo" and "foo", options that exist
only with "debug-foo" and options that exist only with "foo", and make
them into combinable attributes that holds common cflags, extra cflags
for debuggin and extra cflags for non-debugging configurations.
The next step is to make it possible to have template configurations.
Emilia Kasper [Sat, 14 Mar 2015 04:10:13 +0000 (21:10 -0700)]
Fix undefined behaviour in shifts.
Td4 and Te4 are arrays of u8. A u8 << int promotes the u8 to an int first then shifts.
If the mathematical result of a shift (as modelled by lhs * 2^{rhs}) is not representable
in an integer, behaviour is undefined. In other words, you can't shift into the sign bit
of a signed integer. Fix this by casting to u32 whenever we're shifting left by 24.
Matt Caswell [Wed, 11 Mar 2015 17:43:38 +0000 (17:43 +0000)]
Fix RSA_X931_derive_ex
In the RSA_X931_derive_ex a call to BN_CTX_new is made. This can return
NULL on error. However the return value is not tested until *after* it is
derefed! Also at the top of the function a test is made to ensure that
|rsa| is not NULL. If it is we go to the "err" label. Unfortunately the
error handling code deref's rsa.
Matt Caswell [Wed, 11 Mar 2015 17:01:38 +0000 (17:01 +0000)]
SSL_check_chain fix
If SSL_check_chain is called with a NULL X509 object or a NULL EVP_PKEY
or the type of the public key is unrecognised then the local variable
|cpk| in tls1_check_chain does not get initialised. Subsequently an
attempt is made to deref it (after the "end" label), and a seg fault will
result.
Reviewed-by: Dr. Stephen Henson <steve@openssl.org>
Matt Caswell [Wed, 11 Mar 2015 20:19:08 +0000 (20:19 +0000)]
Fix dsa_pub_encode
The return value from ASN1_STRING_new() was not being checked which could
lead to a NULL deref in the event of a malloc failure. Also fixed a mem
leak in the error path.
Matt Caswell [Wed, 11 Mar 2015 20:08:16 +0000 (20:08 +0000)]
Fix dh_pub_encode
The return value from ASN1_STRING_new() was not being checked which could
lead to a NULL deref in the event of a malloc failure. Also fixed a mem
leak in the error path.
Matt Caswell [Wed, 11 Mar 2015 16:00:01 +0000 (16:00 +0000)]
ASN1_primitive_new NULL param handling
ASN1_primitive_new takes an ASN1_ITEM * param |it|. There are a couple
of conditional code paths that check whether |it| is NULL or not - but
later |it| is deref'd unconditionally. If |it| was ever really NULL then
this would seg fault. In practice ASN1_primitive_new is marked as an
internal function in the public header file. The only places it is ever
used internally always pass a non NULL parameter for |it|. Therefore, change
the code to sanity check that |it| is not NULL, and remove the conditional
checking.
Reviewed-by: Rich Salz <rsalz@openssl.org> Reviewed-by: Andy Polyakov <appro@openssl.org>