Tomas Kuthan [Wed, 30 Sep 2015 13:18:05 +0000 (15:18 +0200)]
Check output params on GSS OID set functions
Add sanity checks for the output parameters of
generic_gss_create_empty_oid_set() and
generic_gss_add_oid_set_member(), which are used directly by the API
functions gss_create_empty_oid_set() and gss_add_oid_set_member().
k5_utf8s_to_ucs2s() reads and ignores one extra byte from the input
string before terminating its loop, possibly overrunning the input
buffer of its caller. This overrun is typically without consequence,
but can show up in tools like asan or valgrind during RC4
string-to-key operations. Fix the bug by swapping the order of the
loop conditions.
Tomas Kuthan [Wed, 16 Sep 2015 10:13:26 +0000 (12:13 +0200)]
Fix error mappings for IOV MIC mechglue funcs
The mechglue functions gss_get_mic_iov(), gss_get_mic_iov_length(),
and gss_verify_mic_iov() don't call map_error() to map
mechanism-specific error codes. As a result, a subsequent call to
gss_display_status() fails with GSS_S_BAD_MECH, because no translation
for the error code is found in the error table.
This patch adds the missing map_error call.
[ghudson@mit.edu: correct a whitespace issue, edit commit message]
Ticket #7223 added new policy fields and a new dump format version to
marshal them, but did not add a new iprop dump format version. As a
result, slave KDCs running 1.11 or later cannot receive full resyncs
from master KDCs running 1.10 or earlier. (Reported by John
Devitofranceschi.)
Retroactively add support for pre-1.11 policy entries by making
process_r1_11_policy() read the first ten fields, check whether the
next whitespace character is a newline, and then read the rest if it
is not.
Greg Hudson [Mon, 2 Nov 2015 03:46:56 +0000 (22:46 -0500)]
Fix SPNEGO context import
The patches for CVE-2015-2695 did not implement a SPNEGO
gss_import_sec_context() function, under the erroneous belief that an
exported SPNEGO context would be tagged with the underlying context
mechanism. Implement it now to allow SPNEGO contexts to be
successfully exported and imported after establishment.
Greg Hudson [Mon, 2 Nov 2015 03:45:21 +0000 (22:45 -0500)]
Fix IAKERB context export/import [CVE-2015-2698]
The patches for CVE-2015-2696 contained a regression in the newly
added IAKERB iakerb_gss_export_sec_context() function, which could
cause it to corrupt memory. Fix the regression by properly
dereferencing the context_handle pointer before casting it.
Also, the patches did not implement an IAKERB gss_import_sec_context()
function, under the erroneous belief that an exported IAKERB context
would be tagged as a krb5 context. Implement it now to allow IAKERB
contexts to be successfully exported and imported after establishment.
CVE-2015-2698:
In any MIT krb5 release with the patches for CVE-2015-2696 applied, an
application which calls gss_export_sec_context() may experience memory
corruption if the context was established using the IAKERB mechanism.
Historically, some vulnerabilities of this nature can be translated
into remote code execution, though the necessary exploits must be
tailored to the individual application and are usually quite
complicated.
In build_principal_va(), use k5memdup0() instead of strdup() to make a
copy of the realm, to ensure that we allocate the correct number of
bytes and do not read past the end of the input string. This bug
affects krb5_build_principal(), krb5_build_principal_va(), and
krb5_build_principal_alloc_va(). krb5_build_principal_ext() is not
affected.
CVE-2015-2697:
In MIT krb5 1.7 and later, an authenticated attacker may be able to
cause a KDC to crash using a TGS request with a large realm field
beginning with a null byte. If the KDC attempts to find a referral to
answer the request, it constructs a principal name for lookup using
krb5_build_principal() with the requested realm. Due to a bug in this
function, the null byte causes only one byte be allocated for the
realm field of the constructed principal, far less than its length.
Subsequent operations on the lookup principal may cause a read beyond
the end of the mapped memory region, causing the KDC process to crash.
Nicolas Williams [Mon, 14 Sep 2015 16:28:36 +0000 (12:28 -0400)]
Fix IAKERB context aliasing bugs [CVE-2015-2696]
The IAKERB mechanism currently replaces its context handle with the
krb5 mechanism handle upon establishment, under the assumption that
most GSS functions are only called after context establishment. This
assumption is incorrect, and can lead to aliasing violations for some
programs. Maintain the IAKERB context structure after context
establishment and add new IAKERB entry points to refer to it with that
type. Add initiate and established flags to the IAKERB context
structure for use in gss_inquire_context() prior to context
establishment.
CVE-2015-2696:
In MIT krb5 1.9 and later, applications which call
gss_inquire_context() on a partially-established IAKERB context can
cause the GSS-API library to read from a pointer using the wrong type,
generally causing a process crash. Java server applications using the
native JGSS provider are vulnerable to this bug. A carefully crafted
IAKERB packet might allow the gss_inquire_context() call to succeed
with attacker-determined results, but applications should not make
access control decisions based on gss_inquire_context() results prior
to context establishment.
Nicolas Williams [Mon, 14 Sep 2015 16:27:52 +0000 (12:27 -0400)]
Fix SPNEGO context aliasing bugs [CVE-2015-2695]
The SPNEGO mechanism currently replaces its context handle with the
mechanism context handle upon establishment, under the assumption that
most GSS functions are only called after context establishment. This
assumption is incorrect, and can lead to aliasing violations for some
programs. Maintain the SPNEGO context structure after context
establishment and refer to it in all GSS methods. Add initiate and
opened flags to the SPNEGO context structure for use in
gss_inquire_context() prior to context establishment.
CVE-2015-2695:
In MIT krb5 1.5 and later, applications which call
gss_inquire_context() on a partially-established SPNEGO context can
cause the GSS-API library to read from a pointer using the wrong type,
generally causing a process crash. This bug may go unnoticed, because
the most common SPNEGO authentication scenario establishes the context
after just one call to gss_accept_sec_context(). Java server
applications using the native JGSS provider are vulnerable to this
bug. A carefully crafted SPNEGO packet might allow the
gss_inquire_context() call to succeed with attacker-determined
results, but applications should not make access control decisions
based on gss_inquire_context() results prior to context establishment.
Although our built-in KDB modules do not support client referrals for
AS requests, the KDC is supposed to return one if a third-party module
returns a DB entry containing a principal in a foreign realm.
Unfortunately, this code has never worked; in prepare_error_as(), we
erroneously compare the protocol code errcode against the com_err code
KRB5KDC_ERR_WRONG_REALM; as a result, we never supply the canonical
client principal. Fix this by comparing errcode against the protocol
code KDC_ERR_WRONG_REALM instead.
In the mechglue gss_export_sec_context(), make sure to delete the
union context if the underlying mech context has been deleted. This
can happen if the mech's gss_export_sec_context() returns a failure
and deletes the context (not a behavior exhibited by any of our
in-tree mechanisms, but an allowed behavior for other mechs), or if we
fail to allocate space for the wrapped token.
[ghudson@mit.edu: commit message; rename exit label to "cleanup" and
make it valid for all exit cases]
Solly Ross [Thu, 27 Aug 2015 19:55:35 +0000 (15:55 -0400)]
Check for null name_type in gss_display_name_ext
It is possible for the input name's name_type to be GSS_C_NO_OID.
g_OID_equal() does not account for GSS_C_NO_OID, so we have to
manually check before use to prevent null pointer dereferences.
Before this patch, libkrad would follow the same exact logic for all
socket types when the retries parameter was non-zero. This meant that
when connecting with SOCK_STREAM, multiple requests were sent in case
of packet drops, which, of course, cannot happen for SOCK_STREAM.
Instead, just disable retries for SOCK_STREAM sockets.
Release any previous value of ctx->err_padata before setting it in
init_creds_step_reply(). It could have a prior value after a realm
referral or retriable error.
crypto_retrieve_cert_sans() is allowed to set its princs output to
NULL, although the OpenSSL implementation rarely does. Fix the
TRACE_PKINIT_CLIENT_SAN_KDCCERT_PRINC for loop to allow this like other
parts of the function do, and also get rid of the unnecessary princptr
variable by using an integer index like other parts of the function.
Greg Hudson [Wed, 10 Jun 2015 23:48:51 +0000 (19:48 -0400)]
Tolerate null oid pointer in gss_release_oid()
Under some circumstances, gss_inquire_name() can call
gss_release_oid() with a null oid pointer, which currently causes a
null dereference. The least invasive fix is for gss_release_oid() to
check for the invalid null pointer and return an error, like other
GSS-API functions do.
Greg Hudson [Tue, 24 Mar 2015 16:02:37 +0000 (12:02 -0400)]
Prevent requires_preauth bypass [CVE-2015-2694]
In the OTP kdcpreauth module, don't set the TKT_FLG_PRE_AUTH bit until
the request is successfully verified. In the PKINIT kdcpreauth
module, don't respond with code 0 on empty input or an unconfigured
realm. Together these bugs could cause the KDC preauth framework to
erroneously treat a request as pre-authenticated.
CVE-2015-2694:
In MIT krb5 1.12 and later, when the KDC is configured with PKINIT
support, an unauthenticated remote attacker can bypass the
requires_preauth flag on a client principal and obtain a ciphertext
encrypted in the principal's long-term key. This ciphertext could be
used to conduct an off-line dictionary attack against the user's
password.
krb5_ldap_get_value() takes a pointer to int, and should not be passed
a pointer to any integral type which might have a different width.
Use an intermediate variable for each call.
The erroneous calls in ldap_misc.c were passing pointers to int32_t,
which is harmless on all common platforms. The calls in
ldap_tkt_policy.c were passing pointers to long; on big-endian LP64
platforms, the result would be written to the high 32 bits of the long
value.
Greg Hudson [Thu, 26 Mar 2015 16:47:06 +0000 (12:47 -0400)]
Disable principal renames for LDAP
The current principal rename procedure does not work with the LDAP KDB
module, instead having the effect of deleting the principal. The fix
is not easy and requires amending the DAL (see issue #8065). For now,
detect LDAP and error out when a rename operation is attempted.
Greg Hudson [Fri, 13 Mar 2015 17:30:49 +0000 (13:30 -0400)]
Log invalid restrictions strings
In kadm5int_acl_parse_restrictions(), output a log message if we break
out of the parsing loop with an error. The current structure of the
loop makes it difficult to pinpoint the bad restrictions field, so
just output the whole string.
Greg Hudson [Fri, 13 Mar 2015 16:45:27 +0000 (12:45 -0400)]
Document correct flag names for kadm5.acl
kadm5.acl entries can include restrictions which can force flag values
on or off. These flag values are parsed with krb5_string_to_flags(),
which means the flag names are the ones for default_principal_flags,
not the ones for kadmin addprinc/modprinc.
Greg Hudson [Thu, 12 Mar 2015 20:36:33 +0000 (16:36 -0400)]
Fix scope of kadmind ACL wildcard back-references
In kadm5int_acl_find_entry(), clear the wildcard back-references list
for each acl entry. Otherwise the wildcards we process can affect
back-references for later entries.
Greg Hudson [Thu, 21 Nov 2013 21:18:27 +0000 (16:18 -0500)]
Add another kadmin ACL test for backreferences
Add a test using backreferences which don't correspond directly to
principal components, to verify that *N refers to the Nth wildcard and
not the Nth component.
Solly Ross [Thu, 5 Mar 2015 18:22:58 +0000 (13:22 -0500)]
Import names immediately with COMPOSITE_EXPORT
RFC 6680 specifies that GSS_Export_name_composite() "outputs a token that
"can be imported with GSS_Import_name(), using GSS_C_NT_COMPOSITE_EXPORT
as the name type...". Therefore, in the gss_import_name mechglue, we
should perform the import process imediately when either
GSS_C_NT_COMPOSITE_EXPORT or GSS_C_NT_EXPORT_NAME are used (not just
for the later, as is the current functionality).
The naming extension test was also updated to display the result
of importing with GSS_C_NT_COMPOSITE_EXPORT in addition to
GSS_C_NT_EXPORT_NAME.
Greg Hudson [Tue, 9 Dec 2014 17:37:44 +0000 (12:37 -0500)]
Fix krb5_read_message handling [CVE-2014-5355]
In recvauth_common, do not use strcmp against the data fields of
krb5_data objects populated by krb5_read_message(), as there is no
guarantee that they are C strings. Instead, create an expected
krb5_data value and use data_eq().
In the sample user-to-user server application, check that the received
client principal name is null-terminated before using it with printf
and krb5_parse_name.
CVE-2014-5355:
In MIT krb5, when a server process uses the krb5_recvauth function, an
unauthenticated remote attacker can cause a NULL dereference by
sending a zero-byte version string, or a read beyond the end of
allocated storage by sending a non-null-terminated version string.
The example user-to-user server application (uuserver) is similarly
vulnerable to a zero-length or non-null-terminated principal name
string.
The krb5_recvauth function reads two version strings from the client
using krb5_read_message(), which produces a krb5_data structure
containing a length and a pointer to an octet sequence. krb5_recvauth
assumes that the data pointer is a valid C string and passes it to
strcmp() to verify the versions. If the client sends an empty octet
sequence, the data pointer will be NULL and strcmp() will dereference
a NULL pointer, causing the process to crash. If the client sends a
non-null-terminated octet sequence, strcmp() will read beyond the end
of the allocated storage, possibly causing the process to crash.
uuserver similarly uses krb5_read_message() to read a client principal
name, and then passes it to printf() and krb5_parse_name() without
verifying that it is a valid C string.
The krb5_recvauth function is used by kpropd and the Kerberized
versions of the BSD rlogin and rsh daemons. These daemons are usually
run out of inetd or in a mode which forks before processing incoming
connections, so a process crash will generally not result in a
complete denial of service.
Ben Kaduk [Wed, 19 Nov 2014 17:04:46 +0000 (12:04 -0500)]
Support keyless principals in LDAP [CVE-2014-5354]
Operations like "kadmin -q 'addprinc -nokey foo'" or
"kadmin -q 'purgekeys -all foo'" result in principal entries with
no keys present, so krb5_encode_krbsecretkey() would just return
NULL, which then got unconditionally dereferenced in
krb5_add_ber_mem_ldap_mod().
Apply some fixes to krb5_encode_krbsecretkey() to handle zero-key
principals better, correct the test for an allocation failure, and
slightly restructure the cleanup handler to be shorter and more
appropriate for the usage. Once it no longer short-circuits when
n_key_data is zero, it will produce an array of length two with both
entries NULL, which is treated as an empty list by the LDAP library,
the correct behavior for a keyless principal.
However, attributes with empty values are only handled by the LDAP
library for Modify operations, not Add operations (which only get
a sequence of Attribute, with no operation field). Therefore, only
add an empty krbprincipalkey to the modlist when we will be performing a
Modify, and not when we will be performing an Add, which is conditional
on the (misspelled) create_standalone_prinicipal boolean.
CVE-2014-5354:
In MIT krb5, when kadmind is configured to use LDAP for the KDC
database, an authenticated remote attacker can cause a NULL
dereference by inserting into the database a principal entry which
contains no long-term keys.
In order for the LDAP KDC backend to translate a principal entry
from the database abstraction layer into the form expected by the
LDAP schema, the principal's keys are encoded into a
NULL-terminated array of length-value entries to be stored in the
LDAP database. However, the subroutine which produced this array
did not correctly handle the case where no keys were present,
returning NULL instead of an empty array, and the array was
unconditionally dereferenced while adding to the list of LDAP
operations to perform.
Versions of MIT krb5 prior to 1.12 did not expose a way for
principal entries to have no long-term key material, and
therefore are not vulnerable.
Greg Hudson [Fri, 5 Dec 2014 19:01:39 +0000 (14:01 -0500)]
Fix LDAP misused policy name crash [CVE-2014-5353]
In krb5_ldap_get_password_policy_from_dn, if LDAP_SEARCH returns
successfully with no results, return KRB5_KDB_NOENTRY instead of
returning success with a zeroed-out policy object. This fixes a null
dereference when an admin attempts to use an LDAP ticket policy name
as a password policy name.
CVE-2014-5353:
In MIT krb5, when kadmind is configured to use LDAP for the KDC
database, an authenticated remote attacker can cause a NULL dereference
by attempting to use a named ticket policy object as a password policy
for a principal. The attacker needs to be authenticated as a user who
has the elevated privilege for setting password policy by adding or
modifying principals.
Queries to LDAP scoped to the krbPwdPolicy object class will correctly
not return entries of other classes, such as ticket policy objects, but
may return success with no returned elements if an object with the
requested DN exists in a different object class. In this case, the
routine to retrieve a password policy returned success with a password
policy object that consisted entirely of zeroed memory. In particular,
accesses to the policy name will dereference a NULL pointer. KDC
operation does not access the policy name field, but most kadmin
operations involving the principal with incorrect password policy
will trigger the crash.
Greg Hudson [Wed, 5 Nov 2014 19:12:35 +0000 (14:12 -0500)]
Fix input race condition in t_skew.py
In two of the kinit tests run by t_skew.py, we expect kinit to exit
before reading the password. If we supply a password input for those
commands, we can fail with a broken pipe exception if the master
process tries to write the password after the slave process exits.
Also correctly check the output of the last kinit invocation.
Greg Hudson [Wed, 29 Oct 2014 16:16:40 +0000 (12:16 -0400)]
Remove length limit on PKINIT PKCS#12 prompt
Long pathnames can trigger the 128-byte prompt length limit in
pkinit_get_certs_pkcs12. Use asprintf instead of snprintf. Also
check the result of the prompter invocation.
Greg Hudson [Mon, 20 Oct 2014 16:52:45 +0000 (12:52 -0400)]
Report output ccache errors getting initial creds
In init_creds_step_reply, if we get an error storing output
credentials, do set ctx->complete (since retrieving creds or times
will work at this point) but don't suppress the error code.
Tom Yu [Thu, 16 Oct 2014 19:40:33 +0000 (15:40 -0400)]
Parse "ktadd -norandkey" in remote kadmin client
The remote kadmin client would not parse the "-norandkey" option to
the ktadd subcommand, terminating option parsing and possibly causing
options to be interpreted as principal names.
Greg Hudson [Wed, 8 Oct 2014 00:22:52 +0000 (20:22 -0400)]
Use gssalloc_malloc for GSS error tokens
In kg_accept_krb5, use gssalloc_malloc when allocating space for the
error token, since it will be freed with gssalloc_free. Using malloc
can cause heap corruption on Windows. This bug was masked by #1445
before 1.12.
Greg Hudson [Wed, 14 Jan 2015 18:10:39 +0000 (13:10 -0500)]
Check for null *iter_p in profile_iterator()
In profile_iterator(), return PROF_MAGIC_ITERATOR if *iter_p is NULL,
instead of dereferencing a null pointer, as we did prior to 1.10.
Correct calling code will not trigger this case, but incorrect code
has been reported in the field.
Simo Sorce [Tue, 20 Jan 2015 18:48:34 +0000 (13:48 -0500)]
Do not loop on principal unknown errors
If the canonicalize flag is set, the MIT KDC always return the client
principal when KRB5_KDC_ERR_C_PRICIPAL_UNKNOWN is returned.
Check that this is really a referral by testing that the returned
client realm differs from the requested one.
[ghudson@mit.edu: simplified and narrowed is_referral() contract.
Note that a WRONG_REALM response with e-data or FAST error padata
could now be passed through k5_preauth_tryagain() if it has an empty
crealm or a crealm equal to the requested client realm. Such a
response is unexpected in practice and there is nothing dangerous
about handling it this way.]
Patrik Kis [Fri, 24 Oct 2014 14:15:05 +0000 (16:15 +0200)]
Add missing salt from enctype in t_kdb.py test
In commit 78a7f2a02b82bf297817cd717f092ead40b575b2, which added a test
for bug #7980, some command invocations used new syntax that omitted
salt types from the keysalt list. This omission of salt types is not
supported on krb5-1.12, resulting in test failures. Update to use the
valid (old) key salt syntax.
Greg Hudson [Mon, 29 Dec 2014 18:17:56 +0000 (13:17 -0500)]
Fix gssrpc data leakage [CVE-2014-9423]
[MITKRB5-SA-2015-001] In svcauth_gss_accept_sec_context(), do not copy
bytes from the union context into the handle field we send to the
client. We do not use this handle field, so just supply a fixed
string of "xxxx".
In gss_union_ctx_id_struct, remove the unused "interposer" field which
was causing part of the union context to remain uninitialized.
Greg Hudson [Mon, 29 Dec 2014 18:27:42 +0000 (13:27 -0500)]
Fix kadmind server validation [CVE-2014-9422]
[MITKRB5-SA-2015-001] In kadmind's check_rpcsec_auth(), use
data_eq_string() instead of strncmp() to check components of the
server principal, so that we don't erroneously match left substrings
of "kadmin", "history", or the realm.
Greg Hudson [Sat, 27 Dec 2014 19:16:13 +0000 (14:16 -0500)]
Fix kadm5/gssrpc XDR double free [CVE-2014-9421]
[MITKRB5-SA-2015-001] In auth_gssapi_unwrap_data(), do not free
partial deserialization results upon failure to deserialize. This
responsibility belongs to the callers, svctcp_getargs() and
svcudp_getargs(); doing it in the unwrap function results in freeing
the results twice.
In xdr_krb5_tl_data() and xdr_krb5_principal(), null out the pointers
we are freeing, as other XDR functions such as xdr_bytes() and
xdr_string().
Tom Yu [Wed, 4 Feb 2015 21:09:16 +0000 (16:09 -0500)]
Fix gss_process_context_token() [CVE-2014-5352]
[MITKRB5-SA-2015-001] The krb5 gss_process_context_token() should not
actually delete the context; that leaves the caller with a dangling
pointer and no way to know that it is invalid. Instead, mark the
context as terminated, and check for terminated contexts in the GSS
functions which expect established contexts. Also add checks in
export_sec_context and pseudo_random, and adjust t_prf.c for the
pseudo_random check.
Ben Kaduk [Thu, 21 Aug 2014 22:56:24 +0000 (18:56 -0400)]
Let libgssapi see TGTs in the MSLSA cache
When the current user is a local administrator of a windows machine
where User Account Control (UAC) is enabled, the Windows LSA will
return a block of zeros as the session key for any TGT entry in the
MSLSA: cache. The lcc_retrieve() implementation checks for such
"null" session keys and prevents them from escaping to callers (as
attempts to use them would encounger strange errors). However,
when the TGT is the only entry in the cache, this filtering prevents
scan_ccache() from detecting that the cache contains non-expired
credentials (and that there is a TGT present).
Since scan_ccache() is only looking at metadata in the cache entries,
and does not need to actually use any tickets or session keys, set
the KRB5_TC_NOTICKET flag on the ccache before scanning it. This
will allow the MSLSA implementation to return a cred for the TGT
entry and cause the GSSAPI credential selection algorithm to function
properly.
Ben Kaduk [Thu, 21 Aug 2014 21:33:11 +0000 (17:33 -0400)]
Add some KDC entries to the registry via WiX
Though our library happily uses DNS, I can't get Windows to
successfully contact KDCs found through the SRV records.
So, we do need to stay in the business of shipping around
KDC entries, after all.
Greg Hudson [Wed, 28 May 2014 22:06:59 +0000 (18:06 -0400)]
Make tcl_kadm5.c work with Tcl 8.6
Directly accessing the result field of Tcl_Interp has been deprecated
for a long time, requires a special define in Tcl 8.6, and will be
impossible in Tcl 9. Use Tcl_SetResult instead. The new error
messages are less helpful than the old ones, but this is just support
infrastructure for old tests, so it isn't important.
Tomas Kuthan [Fri, 1 Aug 2014 13:25:50 +0000 (15:25 +0200)]
Fix LDAP key data segmentation [CVE-2014-4345]
For principal entries having keys with multiple kvnos (due to use of
-keepold), the LDAP KDB module makes an attempt to store all the keys
having the same kvno into a single krbPrincipalKey attribute value.
There is a fencepost error in the loop, causing currkvno to be set to
the just-processed value instead of the next kvno. As a result, the
second and all following groups of multiple keys by kvno are each
stored in two krbPrincipalKey attribute values. Fix the loop to use
the correct kvno value.
CVE-2014-4345:
In MIT krb5, when kadmind is configured to use LDAP for the KDC
database, an authenticated remote attacker can cause it to perform an
out-of-bounds write (buffer overrun) by performing multiple cpw
-keepold operations. An off-by-one error while copying key
information to the new database entry results in keys sharing a common
kvno being written to different array buckets, in an array whose size
is determined by the number of kvnos present. After sufficient
iterations, the extra writes extend past the end of the
(NULL-terminated) array. The NULL terminator is always written after
the end of the loop, so no out-of-bounds data is read, it is only
written.
Historically, it has been possible to convert an out-of-bounds write
into remote code execution in some cases, though the necessary
exploits must be tailored to the individual application and are
usually quite complicated. Depending on the allocated length of the
array, an out-of-bounds write may also cause a segmentation fault
and/or application crash.
profile_rename_section should demand only one name.
profile_add_relation should demand only one name if it is creating a
new section. It aso needs to reset state before calling
profile_find_node for the section, in case it didn't look up any
parent sections previously.
In profile_find_node, skip deleted nodes when finding the second
match. Otherwise, profile_clear_nodes could return an error if a node
has some values to clear but the last one is deleted.
In profile_node_iterator, skip deleted nodes when looking up the
section names. Otherwise we could iterate over a deleted section
and/or ignore its replacement.
Greg Hudson [Tue, 18 Feb 2014 06:14:01 +0000 (01:14 -0500)]
Use system dictionary for db2 tests again
The built-in word list is not long enough for all of the libdb2 tests
to run properly. Revert d21a86e47a7cda29225013e08d060095b94b2ee7 and
go back to using the system dictionary if we find one. However, omit
any lines from the chosen word list which contain non-alphabetical
characters.
Greg Hudson [Mon, 17 Feb 2014 05:18:41 +0000 (00:18 -0500)]
Use TAILQ macros instead of CIRCLEQ in libdb2
The optimizer in gcc 4.8.1 (but not the current gcc head revision)
breaks the queue.h CIRCLEQ macros, apparently due to an overzealous
strict aliasing deduction. Use TAILQ macros in the libdb2 mpool code
instead.
Greg Hudson [Thu, 12 Jun 2014 18:34:26 +0000 (14:34 -0400)]
Remove indent workaround in man page RST sources
docutils 0.10 properly adds indentation to example blocks in man
pages, so we do not need to force an extra indentation level. Get rid
of the workaround wherever we use it.
When processing a continuation token, acc_ctx_cont was dereferencing
the initial byte of the token without checking the length. This could
result in a null dereference.
CVE-2014-4344:
In MIT krb5 1.5 and newer, an unauthenticated or partially
authenticated remote attacker can cause a NULL dereference and
application crash during a SPNEGO negotiation by sending an empty
token as the second or later context token from initiator to acceptor.
The attacker must provide at least one valid context token in the
security context negotiation before sending the empty token. This can
be done by an unauthenticated attacker by forcing SPNEGO to
renegotiate the underlying mechanism, or by using IAKERB to wrap an
unauthenticated AS-REQ as the first token.
David Woodhouse [Tue, 15 Jul 2014 16:54:15 +0000 (12:54 -0400)]
Fix double-free in SPNEGO [CVE-2014-4343]
In commit cd7d6b08 ("Verify acceptor's mech in SPNEGO initiator") the
pointer sc->internal_mech became an alias into sc->mech_set->elements,
which should be considered constant for the duration of the SPNEGO
context. So don't free it.
CVE-2014-4343:
In MIT krb5 releases 1.10 and newer, an unauthenticated remote
attacker with the ability to spoof packets appearing to be from a
GSSAPI acceptor can cause a double-free condition in GSSAPI initiators
(clients) which are using the SPNEGO mechanism, by returning a
different underlying mechanism than was proposed by the initiator. At
this stage of the negotiation, the acceptor is unauthenticated, and
the acceptor's response could be spoofed by an attacker with the
ability to inject traffic to the initiator.
Historically, some double-free vulnerabilities can be translated into
remote code execution, though the necessary exploits must be tailored
to the individual application and are usually quite
complicated. Double-frees can also be exploited to cause an
application crash, for a denial of service. However, most GSSAPI
client applications are not vulnerable, as the SPNEGO mechanism is not
used by default (when GSS_C_NO_OID is passed as the mech_type argument
to gss_init_sec_context()). The most common use of SPNEGO is for
HTTP-Negotiate, used in web browsers and other web clients. Most such
clients are believed to not offer HTTP-Negotiate by default, instead
requiring a whitelist of sites for which it may be used to be
configured. If the whitelist is configured to only allow
HTTP-Negotiate over TLS connections ("https://"), a successful
attacker must also spoof the web server's SSL certificate, due to the
way the WWW-Authenticate header is sent in a 401 (Unauthorized)
response message. Unfortunately, many instructions for enabling
HTTP-Negotiate in common web browsers do not include a TLS
requirement.
Greg Hudson [Thu, 17 Oct 2013 18:02:14 +0000 (14:02 -0400)]
Fix race in util/profile/Makefile.in
$(BUILDTOP)/include/profile.h was being updated by two different
rules, which could collide with make -j. Use a dependency from
includes instead of a redundant rule.
Greg Hudson [Wed, 11 Jun 2014 03:53:31 +0000 (23:53 -0400)]
Fix several memory leaks in LDAP KDB modules
Fix memory leaks discovered by running valgrind over kdbtest, and some
related leaks. Many of them result from not calling ldap_msgfree
after an unsuccessful search (as the OpenLDAP documentation requires)
or after an exception following a search, so many of the fixes move or
add ldap_msgfree calls to cleanup labels.
ldap_osa_free_princ_ent was not used, and could not be used because it
frees the container while krb5_lookup_tl_kadm_data uses a
caller-allocated container. Change it to leave the container alone,
but to correctly destroy xdrs. Use it in krb5_ldap_put_principal
where princ_ent was leaked.
In krb5_ldap_put_principal, subtreelist is declared twice in interior
scopes and not properly freed; move it to function scope and free it
up in the cleanup label. Also in krb5_ldap_put_principal, avoiding
decoding multiple KBR5_TL_KADM_DATA values (which we don't expect to
see) as later decodes would cause earlier decodes to leak.
In krb5_encode_krbsecretkey, fix a leak of the krb5_data container and
also add an error check when calling asn1_encode_sequence_of_keys;
otherwise we would dereference a null pointer if we run out of memory
encoding keys (very unlikely).
Ben Kaduk [Thu, 3 Jul 2014 14:42:21 +0000 (10:42 -0400)]
Fix build on systems without RTM_OLD*
For example, FreeBSD has removed RTM_OLDADD and RTM_OLDDEL from its API
in March 2014, with the message:
Garbage collect long time obsoleted (or never used) stuff from routing API
Only attempt to define behavior for these cases if they are defined.
Neng Xue [Mon, 30 Jun 2014 21:04:56 +0000 (14:04 -0700)]
Fix unlikely null dereference in TGS client code
If krb5_get_tgs_ktypes fails (due to an out-of-memory condition or an
error re-reading the profile), k5_make_tgs_req will dereference a null
pointer. Check the return value before dereferencing defenctypes.
Greg Hudson [Wed, 18 Jun 2014 16:58:39 +0000 (12:58 -0400)]
Fix KDC worker process argument parsing
To create worker processes, the KDC shuts down realms, forks off the
worker processes, then reinitializes realms in each child.
Reinitializing realms requires making a second pass over the
command-line arguments. To do this with getopt, optind must be
reinitialized to 1 for each pass; otherwise, no options will be seen
the second time around.
Greg Hudson [Thu, 19 Jun 2014 17:49:16 +0000 (13:49 -0400)]
Handle invalid RFC 1964 tokens [CVE-2014-4341...]
Detect the following cases which would otherwise cause invalid memory
accesses and/or integer underflow:
* An RFC 1964 token being processed by an RFC 4121-only context
[CVE-2014-4342]
* A header with fewer than 22 bytes after the token ID or an
incomplete checksum [CVE-2014-4341 CVE-2014-4342]
* A ciphertext shorter than the confounder [CVE-2014-4341]
* A declared padding length longer than the plaintext [CVE-2014-4341]
If we detect a bad pad byte, continue on to compute the checksum to
avoid creating a padding oracle, but treat the checksum as invalid
even if it compares equal.
CVE-2014-4341:
In MIT krb5, an unauthenticated remote attacker with the ability to
inject packets into a legitimately established GSSAPI application
session can cause a program crash due to invalid memory references
when attempting to read beyond the end of a buffer.
In MIT krb5 releases krb5-1.7 and later, an unauthenticated remote
attacker with the ability to inject packets into a legitimately
established GSSAPI application session can cause a program crash due
to invalid memory references when reading beyond the end of a buffer
or by causing a null pointer dereference.
Nalin Dahyabhai [Wed, 25 Jun 2014 16:56:42 +0000 (12:56 -0400)]
Fix unlikely null dereference in mk_cred()
If krb5_encrypt_keyhelper() returns an error, the ciphertext structure
may contain a non-zero length, but it will already have freed the
pointer to its data, making encrypt_credencpart()'s subsequent attempt
to clear and free the memory fail. Remove that logic.
projects / thirdparty / krb5.git / log
