On Linux, verify that the kernel entropy pool is capable of providing
strong random data before allowing WPA/WPA2 connection to be
established. If 20 bytes of data cannot be read from /dev/random,
force first two 4-way handshakes to fail while collecting entropy
into the internal pool in hostapd. After that, give up on /dev/random
and allow the AP to function based on the combination of /dev/urandom
and whatever data has been collected into the internal entropy pool.
By default, make hostapd and wpa_supplicant maintain an internal
entropy pool that is fed with following information:
hostapd:
- Probe Request frames (timing, RSSI)
- Association events (timing)
- SNonce from Supplicants
wpa_supplicant:
- Scan results (timing, signal/noise)
- Association events (timing)
The internal pool is used to augment the random numbers generated
with the OS mechanism (os_get_random()). While the internal
implementation is not expected to be very strong due to limited
amount of generic (non-platform specific) information to feed the
pool, this may strengthen key derivation on some devices that are
not configured to provide strong random numbers through
os_get_random() (e.g., /dev/urandom on Linux/BSD).
This new mechanism is not supposed to replace proper OS provided
random number generation mechanism. The OS mechanism needs to be
initialized properly (e.g., hw random number generator,
maintaining entropy pool over reboots, etc.) for any of the
security assumptions to hold.
If the os_get_random() is known to provide strong ramdom data (e.g., on
Linux/BSD, the board in question is known to have reliable source of
random data from /dev/urandom), the internal hostapd random pool can be
disabled. This will save some in binary size and CPU use. However, this
should only be considered for builds that are known to be used on
devices that meet the requirements described above. The internal pool
is disabled by adding CONFIG_NO_RANDOM_POOL=y to the .config file.
This commit adds a new wrapper, random_get_bytes(), that is currently
defined to use os_get_random() as is. The places using
random_get_bytes() depend on the returned value being strong random
number, i.e., something that is infeasible for external device to
figure out. These values are used either directly as a key or as
nonces/challenges that are used as input for key derivation or
authentication.
The remaining direct uses of os_get_random() do not need as strong
random numbers to function correctly.
The length of the prime was used incorrectly and this resulted
in WPS DH operation failing whenever the public key ended up having
leading zeros (i.e., about every 1/256th time).
There are no subdirectories in any of these directories or plans
for adding ones. As such, there is no point in running the loop
that does not do anything and can cause problems with some shells.
The returned buffer length was hardcoded to be the prime length
which resulted in shorter results being padded in the end. However,
the results from DH code are supposed to be unpadded (and when used
with WPS, the padding is done in WPS code and it is added to the
beginning of the buffer). This fixes WPS key derivation errors
in about 1/256 of runs ("WPS: Incorrect Authenticator") when using
the internal crypto code.
This allows external programs (e.g., UI) to get more information
about server certificate chain used during TLS handshake. This can
be used both to automatically probe the authentication server to
figure out most likely network configuration and to get information
about reasons for failed authentications.
The follow new control interface events are used for this:
CTRL-EVENT-EAP-PEER-CERT
CTRL-EVENT-EAP-TLS-CERT-ERROR
In addition, there is now an option for matching the server certificate
instead of the full certificate chain for cases where a trusted CA is
not configured or even known. This can be used, e.g., by first probing
the network and learning the server certificate hash based on the new
events and then adding a network configuration with the server
certificate hash after user have accepted it. Future connections will
then be allowed as long as the same server certificate is used.
Authentication server probing can be done, e.g., with following
configuration options:
eap=TTLS PEAP TLS
identity=""
ca_cert="probe://"
Example set of control events for this:
CTRL-EVENT-EAP-STARTED EAP authentication started
CTRL-EVENT-EAP-PROPOSED-METHOD vendor=0 method=21
CTRL-EVENT-EAP-METHOD EAP vendor 0 method 21 (TTLS) selected
CTRL-EVENT-EAP-PEER-CERT depth=0 subject='/C=US/ST=California/L=San Francisco/CN=Server/emailAddress=server@kir.nu' hash=5a1bc1296205e6fdbe3979728efe3920798885c1c4590b5f90f43222d239ca6a
CTRL-EVENT-EAP-TLS-CERT-ERROR reason=8 depth=0 subject='/C=US/ST=California/L=San Francisco/CN=Server/emailAddress=server@kir.nu' err='Server certificate chain probe'
CTRL-EVENT-EAP-FAILURE EAP authentication failed
Server certificate matching is configured with ca_cert, e.g.:
ca_cert="hash://server/sha256/5a1bc1296205e6fdbe3979728efe3920798885c1c4590b5f90f43222d239ca6a"
This functionality is currently available only with OpenSSL. Other
TLS libraries (including internal implementation) may be added in
the future.
Undocumented (at least for the time being) TLS parameters can now
be provided in wpa_supplicant configuration to enable some workarounds
for being able to connect insecurely to some networks. phase1 and
phase2 network parameters can use following options:
tls_allow_md5=1
- allow MD5 signature to be used (disabled by default with GnuTLS)
tls_disable_time_checks=1
- ignore certificate expiration time
For now, only the GnuTLS TLS wrapper implements support for these.
This converts tls_connection_handshake(),
tls_connection_server_handshake(), tls_connection_encrypt(), and
tls_connection_decrypt() to use struct wpa_buf to allow higher layer
code to be cleaned up with consistent struct wpabuf use.
This message from tls_connection_handshake() is not really an error in
most cases, so do not show it if there was indeed no Application Data
available (which is a normal scenario and not an indication of any
error).
This allows libeap.a and libeap.so to be built by merging in multiple
libraries from src subdirectories. In addition, this avoids wasting
extra space and time for local builds.
The following defines are not really needed in most places, so
remove them to clean up source code and build scripts:
EAP_TLS_FUNCS
EAP_TLS_OPENSSL
EAP_TLS_GNUTLS
CONFIG_TLS_INTERNAL
Mainly, this is including header files to get definitions for functions
which is good to verify that the parameters match. None of these are
issues that would have shown as incorrect behavior of the program.
The current MinGW/w32api versions seem to provide all the needed CryptoAPI
functions, so the code for loading these dynamically from the DLL can be
removed.
Need to define the workspace buffer properly to allow compiler to handle
strict aliasing between the incoming unsigned char[64] buffer as an u32
array. The previous version built with strict aliasing enabled can
result in SHA-1 producing incorrect results and consequently, with
4-way handshake failing.
This is based on a report and patch from Dan Williams <dcbw@redhat.com>
but with a different type (the union) used as a fix to avoid needing
extra type casting.
Discovered as part of the investigation of:
https://bugzilla.redhat.com/show_bug.cgi?id=494262#c32
if sha1 is built with gcc without turning off strict aliasing, it will
fail to correctly generate the hashes and will fail its own testcases as
well.
Signed-off-by: Dan Williams <dcbw@redhat.com>
This functionality fits better with src/tls (i.e., internal TLS
implementation), so move it there to make crypto_internal.c more
of a wrapper like other crypto_*.c files.
Private keys can now be used in either unencrypted or encrypted
PKCS #8 encoding. Only the pbeWithMD5AndDES-CBC algorithm (PKCS #5)
is currently supported.
OpenSSL 0.9.7 does not include get_rfc3526_prime_1536() function, so
provide that functionality internally if needed. In addition, make
sha256_vector() building depend on whether SHA256 support is included
in the OpenSSL library. This with CONFIG_INTERNAL_SHA256=y in .config
allows OpenSSL without SHA256 support to be used.
Reorganize the TLS/crypto library segments into a single set of blocks
for each library instead of multiple locations handling library-specific
operations. Group crypto functionality together and get wpa_supplicant
and hostapd Makefile closer to eachother in order to make it easier to
eventually move this into a shared makefile.
Crypto library wrappers can now override the internal DH (group 5)
implementation. As a starting point, this is done with OpenSSL. The
new mechanism is currently available only for WPS (i.e., IKEv2 still
depends on the internal DH implementation).
This allows NSS to be used to derive EAP-TLS/PEAP/TTLS keying material.
NSS requires a patch from
https://bugzilla.mozilla.org/show_bug.cgi?id=507359
to provide the new API. In addition, that patch needs to be modified to
add the 16-bit context length value in SSL_ExportKeyingMaterial() only if
contextlen != 0 in order to match with the EAP-TLS/PEAP/TTLS use cases.
This issue seems to be coming from the unfortunate incompatibility in
draft-ietf-tls-extractor-07.txt (draft-ietf-tls-extractor-00.txt would
have used compatible PRF construction).
At this point, it is unclear how this will be resolved eventually, but
anyway, this shows a mechanism that can be used to implement EAP key
derivation with NSS with a small patch to NSS.
This brings in the first step in adding support for using NSS
(Mozilla Network Security Services) as the crypto and TLS library
with wpa_supplicant. This version is able to run through EAP-PEAP
and EAP-TTLS authentication, but does not yet implement any
certificate/private key configuration. In addition, this does not
implement proper key fetching functions either, so the end result
is not really of much use in real world yet.
The BLOCK_SIZE define can be made more specific by using AES_ prefix and
by moving it to aes.h. After this, most aes-*.c do not really need to
include anything from the internal aes_i.h header file. In other words,
aes_i.h can now be used only for the code that uses the internal AES
block operation implementation and none of the code that can use AES
implementation from an external library do not need to include this
header file.
Better not specify EVP_CIPHER again for the second init call since that
will override key length with the default value. The previous version
was likely to work since most use cases would be likely to use the
default key length. Anyway, better make this handle variable length
ciphers (mainly, RC4), too, just in case it is needed in the future.
This is not really of that much use since rc4_skip() can be used as
easily. In addition, rc4 has caused some symbol conflicts in the past,
so it is easier to live without that as an exported symbol.
wpa_supplicant can now be built with FIPS capable OpenSSL for FIPS mode
operation. Currently, this is only enabling the FIPS mode in OpenSSL
without providing any higher level enforcement in wpa_supplicant.
Consequently, invalid configuration will fail during the authentication
run. Proper configuration (e.g., WPA2-Enterprise with EAP-TLS) allows
the connection to be completed.
Instead of using low level, digest-specific functions, use the generic
EVP interface for digest functions. In addition, report OpenSSL errors
in more detail.
Some crypto libraries can return in these functions (e.g., if a specific
hash function is disabled), so we better provide the caller a chance to
check whether the call failed. The return values are not yet used
anywhere, but they will be needed for future changes.
This removes need for local configuration to ignore *.o and *~
and allows the src/*/.gitignore files to be removed (subdirectories
will inherit the rules from the root .gitignore).
It looks like GnuTLS (at least newer versions) is using random padding
on the application data and the previously used 100 byte extra buffer
for tls_connection_encrypt() calls was not enough to handle all cases.
This resulted in semi-random authentication failures with EAP-PEAP and
EAP-TTLS during Phase 2.
Increase the extra space for encryption from 100 to 300 bytes and add an
error message into tls_gnutls.c to make it easier to notice this issue
should it ever show up again even with the larger buffer.
Updated OpenSSL code for EAP-FAST to use an updated version of the
session ticket overriding API that was included into the upstream
OpenSSL 0.9.9 tree on 2008-11-15 (no additional OpenSSL patch is
needed with that version anymore).
When the TLS handshake had been completed earlier by the server in case of
abbreviated handshake, the output buffer length was left uninitialized. It
must be initialized to zero in this case. This code is used by EAP-FAST
server and the uninitialized length could have caused it to try to send a
very large frame (though, this would be terminated by the 50 roundtrip EAP
limit). This broke EAP-FAST server code in some cases when PAC was used to
establish the tunnel.
The internal TLS implementation can now use both PKCS #1 RSA private key
and PKCS #8 encapsulated RSA private key. PKCS #8 encrypted private key is
not yet supported.
The server handshake processing was still using SSL_read() to get OpenSSL
to perform the handshake. While this works for most cases, it caused some
issues for re-authentication. This is now changed to use SSL_accept() which
is more approriate here since we know that the handshake is still going on
and there will not be any tunneled data available. This resolves some of
the re-authentication issues and makes it possible for the server to notice
if TLS processing fails (SSL_read() did not return an error in many of
these cases while SSL_accept() does).
Set session id context to a unique value in order to avoid fatal errors
when client tries session resumption (SSL_set_session_id_context() must be
called for that to work), but disable session resumption with the unique
value for the time being since not all server side code is ready for it yet
(e.g., EAP-TTLS needs special Phase 2 processing when using abbreviated
handshake).
Changed EAP-TLS server not to call TLS library when processing the final
ACK (empty data) from the client in order to avoid starting a new TLS
handshake with SSL_accept().
It looks like this SSL_set_options() value was added in 0.9.9 and it does
not exist in stable releases of OpenSSL. Fix build by using #ifdef on this
variable before use.
The middle byte of the secret (key for PRF) is shared with key halfs in
case the key length is odd. This does not happen in any of the current
tls_prf() uses, but it's better to fix this function to avoid future issues
should someone end up defining a use that uses an odd length for the key.
