This reduces differences in timing and memory access within the
hunting-and-pecking loop for ECC groups that have a prime that is not
close to a power of two (e.g., Brainpool curves).
Signed-off-by: Jouni Malinen <j@w1.fi>
Start sharing common SAE and EAP-pwd functionality by adding a new
source code file that can be included into both. This first step is
bringing in a shared function to check whether a group is suitable.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
Move the identical function used by both SAE and EAP-pwd to
src/utils/common.c to avoid duplicated implementation.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
This reduces timing and memory access pattern differences for an
operation that could depend on the used password.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
None of the ECC groups supported in the implementation had a cofactor
greater than 1, so these checks are unreachable and for all cases, the
cofactor is known to be 1. Furthermore, RFC 5931 explicitly disallow use
of ECC groups with cofactor larger than 1, so this checks cannot be
needed for any curve that is compliant with the RFC.
Remove the unneeded group cofactor checks to simplify the
implementation.
Signed-off-by: Jouni Malinen <j@w1.fi>
Based on the SAE implementation guidance update to not allow ECC groups
with a prime that is under 256 bits, reject groups 25, 26, and 27 in
EAP-pwd.
Signed-off-by: Jouni Malinen <j@w1.fi>
RFC 5931 has these conditions as MUST requirements, so better follow
them explicitly even if the rand,mask == 0 or rand+mask == 0 or 1 cases
are very unlikely to occur in practice while generating random values
locally.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
This adds an explicit check for 0 < x,y < prime based on RFC 5931,
2.8.5.2.2 requirement. The earlier checks might have covered this
implicitly, but it is safer to avoid any dependency on implicit checks
and specific crypto library behavior. (CVE-2019-9498 and CVE-2019-9499)
Furthermore, this moves the EAP-pwd element and scalar parsing and
validation steps into shared helper functions so that there is no need
to maintain two separate copies of this common functionality between the
server and peer implementations.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
This algorithm could leak information to external observers in form of
timing differences or memory access patterns (cache use). While the
previous implementation had protection against the most visible timing
differences (looping 40 rounds and masking the legendre operation), it
did not protect against memory access patterns between the two possible
code paths in the masking operations. That might be sufficient to allow
an unprivileged process running on the same device to be able to
determine which path is being executed through a cache attack and based
on that, determine information about the used password.
Convert the PWE finding loop to use constant time functions and
identical memory access path without different branches for the QR/QNR
cases to minimize possible side-channel information similarly to the
changes done for SAE authentication. (CVE-2019-9495)
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
tmp2 (y^2) was derived once in each iteration of the loop and only freed
after all the loop iterations. Fix this by freeing the temporary value
during each iteration.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
Run through the hunting-and-pecking loop 40 times to mask the time
necessary to find PWE. The odds of PWE not being found in 40 loops is
roughly 1 in 1 trillion.
Signed-off-by: Dan Harkins <dharkins@lounge.org>
BoringSSL is Google's cleanup of OpenSSL and an attempt to unify
Chromium, Android and internal codebases around a single OpenSSL.
As part of moving Android to BoringSSL, the wpa_supplicant maintainers
in Android requested that I upstream the change. I've worked to reduce
the size of the patch a lot but I'm afraid that it still contains a
number of #ifdefs.
[1] https://www.imperialviolet.org/2014/06/20/boringssl.html
Signed-off-by: Adam Langley <agl@chromium.org>
This changes OpenSSL calls to explicitly clear the EC_POINT memory
allocations when freeing them. This adds an extra layer of security by
avoiding leaving potentially private keys into local memory after they
are not needed anymore. While some of these variables are not really
private (e.g., they are sent in clear anyway), the extra cost of
clearing them is not significant and it is simpler to just clear these
explicitly rather than review each possible code path to confirm where
this does not help.
Signed-off-by: Florent Daigniere <nextgens@freenetproject.org>
This changes OpenSSL calls to explicitly clear the bignum memory
allocations when freeing them. This adds an extra layer of security by
avoiding leaving potentially private keys into local memory after they
are not needed anymore. While some of these variables are not really
private (e.g., they are sent in clear anyway), the extra cost of
clearing them is not significant and it is simpler to just clear these
explicitly rather than review each possible code path to confirm where
this does not help.
Signed-off-by: Florent Daigniere <nextgens@freenetproject.org>
At least some error paths (e.g., hitting the limit on hunt-and-peck
iterations) could have resulted in double-freeing of some memory
allocations. Avoid this by setting the pointers to NULL after they have
been freed instead of trying to free the data structure in a location
where some external references cannot be cleared. [Bug 453]
Signed-hostap: Jouni Malinen <j@w1.fi>
The previously used limit (10) is too small for practical purposes since
it can result in about 1 out of 1000 authentication attempts failing.
Increase the limit to 30 to avoid such issues. [Bug 453]
Signed-hostap: Jouni Malinen <j@w1.fi>
intended-for: hostap-1
Remove the GPL notification text from EAP-pwd implementation per
approval from Dan Harkins who contributed these files.
(email from Dan Harkins <dharkins@lounge.org> dated
Wed, 4 Jan 2012 16:25:48 -0800)
Signed-hostap: Jouni Malinen <j@w1.fi>
Another niceness of OpenSSL is that if the high-order bit of a 521-bit
big num is not set then BN_bn2bin() will just return 65 bytes instead of
66 bytes with the 1st (big endian, after all) being all zero. When this
happens the wrong number of octets are mixed into function H(). So
there's a whole bunch of "offset" computations and BN_bn2bin() dumps the
big number into a buffer + offset. That should be obvious in the patch
too.
The previous EAP-pwd KDF implemented has an issue with group 21, that is
an elliptic curve group based on a 521 bit prime. 521 is not an even
multiple of 8, and therein lies the problem.
OpenSSL's BN library interprets a string of bits as in big-endian format
so all the calls of BN_bin2bn() will take the binary blob of bits and
turn it into a big number in big-endian format. In the EAP-pwd KDF, I am
stretching the key to "primebitlen". When that is not an even multiple
of 8 I have to mask off the excess. But I was masking off the excess
bits in the 1st octet (big endian after all) but that isn't right. The
KDF produces a string of endian-less bits. The 521st bit is the first
bit in the last octet, not the 7th bit in the first octet. So that has
been fixed and you can see in the attached diff what I'm doing.
The changes are:
1. the word "and" in the hunting-and-pecking string passed to the KDF
should be capitalized.
2. the primebitlen used in the KDF should be a short not an int.
3. the computation of MK in hostap is based on an older version of the
draft and is not the way it's specified in the RFC.
4. the group being passed into computation of the Commit was not in
network order.
Jouni Malinen