We have always a carry in %rcx or %rbx in range 0..2
from the previous stage, that is added to the result
of the 64-bit square, but the low nibble of any square
can only be 0, 1, 4, 9.
Therefore one "adcq $0, %rdx" can be removed.
Likewise in the ADX code we can remove one
"adcx %rbp, $out" since %rbp is always 0, and carry is
also zero, therefore that is a no-op.
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/10576)
There is an overflow bug in the x64_64 Montgomery squaring procedure used in
exponentiation with 512-bit moduli. No EC algorithms are affected. Analysis
suggests that attacks against 2-prime RSA1024, 3-prime RSA1536, and DSA1024 as a
result of this defect would be very difficult to perform and are not believed
likely. Attacks against DH512 are considered just feasible. However, for an
attack the target would have to re-use the DH512 private key, which is not
recommended anyway. Also applications directly using the low level API
BN_mod_exp may be affected if they use BN_FLG_CONSTTIME.
CVE-2019-1551
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
(Merged from https://github.com/openssl/openssl/pull/10576)
An unintended consequence of https://github.com/openssl/openssl/pull/9808
is that when an explicit parameters curve is matched against one of the
well-known builtin curves we automatically inherit also the associated
seed parameter, even if the input parameters excluded such parameter.
This later affects the serialization of such parsed keys, causing their
input DER encoding and output DER encoding to differ due to the
additional optional field.
This does not cause problems internally but could affect external
applications, as reported in
https://github.com/openssl/openssl/pull/9811#issuecomment-536153288
This commit fixes the issue by conditionally clearing the seed field if
the original input parameters did not include it.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Tomas Mraz <tmraz@fedoraproject.org>
(Merged from https://github.com/openssl/openssl/pull/10141)
An attack is simple, if the first CMS_recipientInfo is valid but the
second CMS_recipientInfo is chosen ciphertext. If the second
recipientInfo decodes to PKCS #1 v1.5 form plaintext, the correct
encryption key will be replaced by garbage, and the message cannot be
decoded, but if the RSA decryption fails, the correct encryption key is
used and the recipient will not notice the attack.
As a work around for this potential attack the length of the decrypted
key must be equal to the cipher default key length, in case the
certifiate is not given and all recipientInfo are tried out.
The old behaviour can be re-enabled in the CMS code by setting the
CMS_DEBUG_DECRYPT flag.
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/9777)
(cherry picked from commit 5840ed0cd1e6487d247efbc1a04136a41d7b3a37)
Description
-----------
Upon `EC_GROUP_new_from_ecparameters()` check if the parameters match any
of the built-in curves. If that is the case, return a new
`EC_GROUP_new_by_curve_name()` object instead of the explicit parameters
`EC_GROUP`.
This affects all users of `EC_GROUP_new_from_ecparameters()`:
- direct calls to `EC_GROUP_new_from_ecparameters()`
- direct calls to `EC_GROUP_new_from_ecpkparameters()` with an explicit
parameters argument
- ASN.1 parsing of explicit parameters keys (as it eventually
ends up calling `EC_GROUP_new_from_ecpkparameters()`)
A parsed explicit parameter key will still be marked with the
`OPENSSL_EC_EXPLICIT_CURVE` ASN.1 flag on load, so, unless
programmatically forced otherwise, if the key is eventually serialized
the output will still be encoded with explicit parameters, even if
internally it is treated as a named curve `EC_GROUP`.
Before this change, creating any `EC_GROUP` object using
`EC_GROUP_new_from_ecparameters()`, yielded an object associated with
the default generic `EC_METHOD`, but this was never guaranteed in the
documentation.
After this commit, users of the library that intentionally want to
create an `EC_GROUP` object using a specific `EC_METHOD` can still
explicitly call `EC_GROUP_new(foo_method)` and then manually set the
curve parameters using `EC_GROUP_set_*()`.
Motivation
----------
This has obvious performance benefits for the built-in curves with
specialized `EC_METHOD`s and subtle but important security benefits:
- the specialized methods have better security hardening than the
generic implementations
- optional fields in the parameter encoding, like the `cofactor`, cannot
be leveraged by an attacker to force execution of the less secure
code-paths for single point scalar multiplication
- in general, this leads to reducing the attack surface
Check the manuscript at https://arxiv.org/abs/1909.01785 for an in depth
analysis of the issues related to this commit.
It should be noted that `libssl` does not allow to negotiate explicit
parameters (as per RFC 8422), so it is not directly affected by the
consequences of using explicit parameters that this commit fixes.
On the other hand, we detected external applications and users in the
wild that use explicit parameters by default (and sometimes using 0 as
the cofactor value, which is technically not a valid value per the
specification, but is tolerated by parsers for wider compatibility given
that the field is optional).
These external users of `libcrypto` are exposed to these vulnerabilities
and their security will benefit from this commit.
Related commits
---------------
While this commit is beneficial for users using built-in curves and
explicit parameters encoding for serialized keys, commit
b783beeadf6b80bc431e6f3230b5d5585c87ef87 (and its equivalents for the
1.0.2, 1.1.0 and 1.1.1 stable branches) fixes the consequences of the
invalid cofactor values more in general also for other curves
(CVE-2019-1547).
The following list covers commits in `master` that are related to the
vulnerabilities presented in the manuscript motivating this commit:
- d2baf88c43 [crypto/rsa] Set the constant-time flag in multi-prime RSA too
- 311e903d84 [crypto/asn1] Fix multiple SCA vulnerabilities during RSA key validation.
- b783beeadf [crypto/ec] for ECC parameters with NULL or zero cofactor, compute it
- 724339ff44 Fix SCA vulnerability when using PVK and MSBLOB key formats
Note that the PRs that contributed the listed commits also include other
commits providing related testing and documentation, in addition to
links to PRs and commits backporting the fixes to the 1.0.2, 1.1.0 and
1.1.1 branches.
This commit includes a partial backport of
https://github.com/openssl/openssl/pull/8555
(commit 8402cd5f75f8c2f60d8bd39775b24b03dd8b3b38)
for which the main author is Shane Lontis.
Responsible Disclosure
----------------------
This and the other issues presented in https://arxiv.org/abs/1909.01785
were reported by Cesar Pereida García, Sohaib ul Hassan, Nicola Tuveri,
Iaroslav Gridin, Alejandro Cabrera Aldaya and Billy Bob Brumley from the
NISEC group at Tampere University, FINLAND.
The OpenSSL Security Team evaluated the security risk for this
vulnerability as low, and encouraged to propose fixes using public Pull
Requests.
_______________________________________________________________________________
Co-authored-by: Shane Lontis <shane.lontis@oracle.com>
(Backport from https://github.com/openssl/openssl/pull/9808)
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/9811)
The cofactor argument to EC_GROUP_set_generator is optional, and SCA
mitigations for ECC currently use it. So the library currently falls
back to very old SCA-vulnerable code if the cofactor is not present.
This PR allows EC_GROUP_set_generator to compute the cofactor for all
curves of cryptographic interest. Steering scalar multiplication to more
SCA-robust code.
This issue affects persisted private keys in explicit parameter form,
where the (optional) cofactor field is zero or absent.
It also affects curves not built-in to the library, but constructed
programatically with explicit parameters, then calling
EC_GROUP_set_generator with a nonsensical value (NULL, zero).
The very old scalar multiplication code is known to be vulnerable to
local uarch attacks, outside of the OpenSSL threat model. New results
suggest the code path is also vulnerable to traditional wall clock
timing attacks.
CVE-2019-1547
Reviewed-by: Nicola Tuveri <nic.tuv@gmail.com>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/9799)
This commit addresses multiple side-channel vulnerabilities present during RSA key validation.
Private key parameters are re-computed using variable-time functions.
This issue was discovered and reported by the NISEC group at TAU Finland.
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/9785)
Replace flip_endian() by using the little endian specific
bn_bn2lebinpad() and bn_lebin2bn().
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
(Merged from https://github.com/openssl/openssl/pull/9793)
This issue was partially addressed by commit
972c87dfc7e765bd28a4964519c362f0d3a58ca4, which hardened its callee
BN_num_bits_word() to avoid leaking the most-significant word of its
argument via branching and memory access pattern.
The commit message also reported:
> There are a few places where BN_num_bits is called on an input where
> the bit length is also secret. This does *not* fully resolve those
> cases as we still only look at the top word.
BN_num_bits() is called directly or indirectly (e.g., through
BN_num_bytes() or BN_bn2binpad() ) in various parts of the `crypto/ec`
code, notably in all the currently supported implementations of scalar
multiplication (in the generic path through ec_scalar_mul_ladder() as
well as in dedicated methods like ecp_nistp{224,256,521}.c and
ecp_nistz256.c).
Under the right conditions, a motivated SCA attacker could retrieve the
secret bitlength of a secret nonce through this vulnerability,
potentially leading, ultimately, to recover a long-term secret key.
With this commit, exclusively for BIGNUMs that are flagged with
BN_FLG_CONSTTIME, instead of accessing only bn->top, all the limbs of
the BIGNUM are accessed up to bn->dmax and bitwise masking is used to
avoid branching.
Memory access pattern still leaks bn->dmax, the size of the lazily
allocated buffer for representing the BIGNUM, which is inevitable with
the current BIGNUM architecture: reading past bn->dmax would be an
out-of-bound read.
As such, it's the caller responsibility to ensure that bn->dmax does not
leak secret information, by explicitly expanding the internal BIGNUM
buffer to a public value sufficient to avoid any lazy reallocation
while manipulating it: this should be already done at the top level
alongside setting the BN_FLG_CONSTTIME.
Thanks to David Schrammel and Samuel Weiser for reporting this issue
through responsible disclosure.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
(Merged from https://github.com/openssl/openssl/pull/9793)
BN_bn2bin() is not constant-time and leaks the number of bits in the
processed BIGNUM.
The specialized methods in ecp_nistp224.c, ecp_nistp256.c and
ecp_nistp521.c internally used BN_bn2bin() to convert scalars into the
internal fixed length representation.
This can leak during ECDSA/ECDH key generation or handling the nonce
while generating an ECDSA signature, when using these implementations.
The amount and risk of leaked information useful for a SCA attack
varies for each of the three curves, as it depends mainly on the
ratio between the bitlength of the curve subgroup order (governing the
size of the secret nonce/key) and the limb size for the internal BIGNUM
representation (which depends on the compilation target architecture).
To fix this, we replace BN_bn2bin() with bn_bn2binpad(), bounding the
output length to the width of the internal representation buffer: this
length is public.
Internally the final implementation of both bn_bn2binpad() and
BN_bn2bin() already has masking in place to avoid leaking bn->top
through memory access patterns.
Memory access pattern still leaks bn->dmax, the size of the lazily
allocated buffer for representing the BIGNUM, which is inevitable with
the current BIGNUM architecture: reading past bn->dmax would be an
out-of-bound read.
As such, it's the caller responsibility to ensure that bn->dmax does not
leak secret information, by explicitly expanding the internal BIGNUM
buffer to a public value sufficient to avoid any lazy reallocation
while manipulating it: this is already done at the top level alongside
setting the BN_FLG_CONSTTIME.
Finally, the internal implementation of bn_bn2binpad() indirectly calls
BN_num_bits() via BN_num_bytes(): the current implementation of
BN_num_bits() can leak information to a SCA attacker, and is addressed
in the next commit.
Thanks to David Schrammel and Samuel Weiser for reporting this issue
through responsible disclosure.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
(Merged from https://github.com/openssl/openssl/pull/9793)
This commit addresses a side-channel vulnerability present when
PVK and MSBLOB key formats are loaded into OpenSSL.
The public key was not computed using a constant-time exponentiation
function.
This issue was discovered and reported by the NISEC group at TAU Finland.
Reviewed-by: Nicola Tuveri <nic.tuv@gmail.com>
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/9638)
The barriers prevent the compiler from narrowing down the
possible value range of the mask and ~mask in the select
statements, which avoids the recognition of the select
and turning it into a conditional load or branch.
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/9419)
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Matthias St. Pierre <Matthias.St.Pierre@ncp-e.com>
(Merged from https://github.com/openssl/openssl/pull/9101)
This is a backport of commit 1e2012b7 to 1.0.2. This hardening change
was made to 1.1.0 but was not backported to 1.0.2. Recent CVEs in user
applications have shown this additional hardening in 1.0.2 would be
beneficial.
E.g. see the patch for CVE-2019-9498
https://w1.fi/security/2019-4/0011-EAP-pwd-server-Verify-received-scalar-and-element.patch
and CVE-2019-9499
https://w1.fi/security/2019-4/0013-EAP-pwd-client-Verify-received-scalar-and-element.patch
The original commit had this description:
We already test in EC_POINT_oct2point that points are on the curve. To
be on the safe side, move this check to
EC_POINT_set_affine_coordinates_* so as to also check point coordinates
received through some other method.
We do not check projective coordinates, though, as
- it's unlikely that applications would be receiving this primarily
internal representation from untrusted sources, and
- it's possible that the projective setters are used in a setting where
performance matters.
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8750)
constant time with a memory access pattern that does not depend
on secret information.
[extended tests]
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/8543)
(cherry picked from commit 9c0cf214e7836eb5aaf1ea5d3cbf6720533f86b5)
The secret point R can be recovered from S using the equation R = S - P.
The X and Z coordinates should be sufficient for that.
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/8505)
(cherry picked from commit 502b871ad4eacc96a31f89d9a9470ca2858da998)
The function felem_diff_128_64 in ecp_nistp521.c substracts the number |in|
from |out| mod p. In order to avoid underflow it first adds 32p mod p
(which is equivalent to 0 mod p) to |out|. The comments and variable naming
suggest that the original author intended to add 64p mod p. In fact it
has been shown that with certain unusual co-ordinates it is possible to
cause an underflow in this function when only adding 32p mod p while
performing a point double operation. By changing this to 64p mod p the
underflow is avoided.
It turns out to be quite difficult to construct points that satisfy the
underflow criteria although this has been done and the underflow
demonstrated. However none of these points are actually on the curve.
Finding points that satisfy the underflow criteria and are also *on* the
curve is considered significantly more difficult. For this reason we do
not believe that this issue is currently practically exploitable and
therefore no CVE has been assigned.
This only impacts builds using the enable-ec_nistp_64_gcc_128 Configure
option.
With thanks to Bo-Yin Yang, Billy Brumley and Dr Liu for their significant
help in investigating this issue.
Reviewed-by: Nicola Tuveri <nic.tuv@gmail.com>
(Merged from https://github.com/openssl/openssl/pull/8405)
(cherry picked from commit 13fbce17fc9f02e2401fc3868f3f8e02d6647e5f)
Thanks to David Benjamin who reported this, performed the analysis and
suggested the patch. I have incorporated some of his analysis in the
comments below.
This issue can cause an out-of-bounds read. It is believed that this was
not reachable until the recent "fixed top" changes. Analysis has so far
only identified one code path that can encounter this - although it is
possible that others may be found. The one code path only impacts 1.0.2 in
certain builds. The fuzzer found a path in RSA where iqmp is too large. If
the input is all zeros, the RSA CRT logic will multiply a padded zero by
iqmp. Two mitigating factors:
- Private keys which trip this are invalid (iqmp is not reduced mod p).
Only systems which take untrusted private keys care.
- In OpenSSL 1.1.x, there is a check which rejects the oversize iqmp,
so the bug is only reproducible in 1.0.2 so far.
Fortunately, the bug appears to be relatively harmless. The consequences of
bn_cmp_word's misbehavior are:
- OpenSSL may crash if the buffers are page-aligned and the previous page is
non-existent.
- OpenSSL will incorrectly treat two BN_ULONG buffers as not equal when they
are equal.
- Side channel concerns.
The first is indeed a concern and is a DoS bug. The second is fine in this
context. bn_cmp_word and bn_cmp_part_words are used to compute abs(a0 - a1)
in Karatsuba. If a0 = a1, it does not matter whether we use a0 - a1 or
a1 - a0. The third would be worth thinking about, but it is overshadowed
by the entire Karatsuba implementation not being constant time.
Due to the difficulty of tripping this and the low impact no CVE is felt
necessary for this issue.
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Viktor Dukhovni <viktor@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8326)
(cherry picked from commit 576129cd72ae054d246221f111aabf42b9c6d76d)
This commit adds a simple unit test to make sure that the constant-time
flag does not "leak" among BN_CTX frames:
- test_ctx_consttime_flag() initializes (and later frees before
returning) a BN_CTX object, then it calls in sequence
test_ctx_set_ct_flag() and test_ctx_check_ct_flag() using the same
BN_CTX object.
- test_ctx_set_ct_flag() starts a frame in the given BN_CTX and sets the
BN_FLG_CONSTTIME flag on some of the BIGNUMs obtained from the frame
before ending it.
- test_ctx_check_ct_flag() then starts a new frame and gets a number of
BIGNUMs from it. In absence of leaks, none of the BIGNUMs in the new
frame should have BN_FLG_CONSTTIME set.
In actual BN_CTX usage inside libcrypto the leak could happen at any
depth level in the BN_CTX stack, with varying results depending on the
patterns of sibling trees of nested function calls sharing the same
BN_CTX object, and the effect of unintended BN_FLG_CONSTTIME on the
called BN_* functions.
This simple unit test abstracts away this complexity and verifies that
the leak does not happen between two sibling functions sharing the same
BN_CTX object at the same level of nesting.
(manually cherry picked from commit fe16ae5f95fa86ddb049a8d1e2caee0b80b32282)
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8295)
Cygwin binaries should not enforce text mode these days, just
use text mode if the underlying mount point requests it
CLA: trivial
Signed-off-by: Corinna Vinschen <vinschen@redhat.com>
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8249)
