This commit leverages the Montgomery ladder scaffold introduced in #6690
(alongside a specialized Lopez-Dahab ladder for binary curves) to
provide a specialized differential addition-and-double implementation to
speedup prime curves, while keeping all the features of
`ec_scalar_mul_ladder` against SCA attacks.
The arithmetic in ladder_pre, ladder_step and ladder_post is auto
generated with tooling, from the following formulae:
- `ladder_pre`: Formula 3 for doubling from Izu-Takagi "A fast parallel
elliptic curve multiplication resistant against side channel attacks",
as described at
https://hyperelliptic.org/EFD/g1p/auto-shortw-xz.html#doubling-dbl-2002-it-2
- `ladder_step`: differential addition-and-doubling Eq. (8) and (10)
from Izu-Takagi "A fast parallel elliptic curve multiplication
resistant against side channel attacks", as described at
https://hyperelliptic.org/EFD/g1p/auto-shortw-xz.html#ladder-ladd-2002-it-3
- `ladder_post`: y-coordinate recovery using Eq. (8) from Brier-Joye
"Weierstrass Elliptic Curves and Side-Channel Attacks", modified to
work in projective coordinates.
Co-authored-by: Nicola Tuveri <nic.tuv@gmail.com>
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6772)
ecp_nistz256_set_from_affine is called when application attempts to use
custom generator, i.e. rarely. Even though it was wrong, it didn't
affect point operations, they were just not as fast as expected.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6738)
The ecp_nistz256_scatter_w7 function is called when application
attempts to use custom generator, i.e. rarely. Even though non-x86_64
versions were wrong, it didn't affect point operations, they were just
not as fast as expected.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6738)
Originally suggested solution for "Return Of the Hidden Number Problem"
is arguably too expensive. While it has marginal impact on slower
curves, none to ~6%, optimized implementations suffer real penalties.
Most notably sign with P-256 went more than 2 times[!] slower. Instead,
just implement constant-time BN_mod_add_quick.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: David Benjamin <davidben@google.com>
(Merged from https://github.com/openssl/openssl/pull/6664)
By default `ec_scalar_mul_ladder` (which uses the Lopez-Dahab ladder
implementation) is used only for (k * Generator) or (k * VariablePoint).
ECDSA verification uses (a * Generator + b * VariablePoint): this commit
forces the use of `ec_scalar_mul_ladder` also for the ECDSA verification
path, while using the default wNAF implementation for any other case.
With this commit `ec_scalar_mul_ladder` loses the static attribute, and
is added to ec_lcl.h so EC_METHODs can directly use it.
While working on a new custom EC_POINTs_mul implementation, I realized
that many checks (e.g. all the points being compatible with the given
EC_GROUP, creating a temporary BN_CTX if `ctx == NULL`, check for the
corner case `scalar == NULL && num == 0`) were duplicated again and
again in every single implementation (and actually some
implementations lacked some of the tests).
I thought that it makes way more sense for those checks that are
independent from the actual implementation and should always be done, to
be moved in the EC_POINTs_mul wrapper: so this commit also includes
these changes.
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6690)
This commit uses the new ladder scaffold to implement a specialized
ladder step based on differential addition-and-doubling in mixed
Lopez-Dahab projective coordinates, modified to independently blind the
operands.
The arithmetic in `ladder_pre`, `ladder_step` and `ladder_post` is
auto generated with tooling:
- see, e.g., "Guide to ECC" Alg 3.40 for reference about the
`ladder_pre` implementation;
- see https://www.hyperelliptic.org/EFD/g12o/auto-code/shortw/xz/ladder/mladd-2003-s.op3
for the differential addition-and-doubling formulas implemented in
`ladder_step`;
- see, e.g., "Fast Multiplication on Elliptic Curves over GF(2**m)
without Precomputation" (Lopez and Dahab, CHES 1999) Appendix Alg Mxy
for the `ladder_post` implementation to recover the `(x,y)` result in
affine coordinates.
Co-authored-by: Billy Brumley <bbrumley@gmail.com>
Co-authored-by: Sohaib ul Hassan <soh.19.hassan@gmail.com>
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6690)
for specialized Montgomery ladder implementations
PR #6009 and #6070 replaced the default EC point multiplication path for
prime and binary curves with a unified Montgomery ladder implementation
with various timing attack defenses (for the common paths when a secret
scalar is feed to the point multiplication).
The newly introduced default implementation directly used
EC_POINT_add/dbl in the main loop.
The scaffolding introduced by this commit allows EC_METHODs to define a
specialized `ladder_step` function to improve performances by taking
advantage of efficient formulas for differential addition-and-doubling
and different coordinate systems.
- `ladder_pre` is executed before the main loop of the ladder: by
default it copies the input point P into S, and doubles it into R.
Specialized implementations could, e.g., use this hook to transition
to different coordinate systems before copying and doubling;
- `ladder_step` is the core of the Montgomery ladder loop: by default it
computes `S := R+S; R := 2R;`, but specific implementations could,
e.g., implement a more efficient formula for differential
addition-and-doubling;
- `ladder_post` is executed after the Montgomery ladder loop: by default
it's a noop, but specialized implementations could, e.g., use this
hook to transition back from the coordinate system used for optimizing
the differential addition-and-doubling or recover the y coordinate of
the result point.
This commit also renames `ec_mul_consttime` to `ec_scalar_mul_ladder`,
as it better corresponds to what this function does: nothing can be
truly said about the constant-timeness of the overall execution of this
function, given that the underlying operations are not necessarily
constant-time themselves.
What this implementation ensures is that the same fixed sequence of
operations is executed for each scalar multiplication (for a given
EC_GROUP), with no dependency on the value of the input scalar.
Co-authored-by: Sohaib ul Hassan <soh.19.hassan@gmail.com>
Co-authored-by: Billy Brumley <bbrumley@gmail.com>
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6690)
Run `make update ERROR_REBUILD=-rebuild` to remove some stale error
codes for SM2 (which is now using its own submodule for error codes,
i.e., `SM2_*`).
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6690)
Move base 2^64 code to own #if section. It was nested in base 2^51 section,
which arguably might have been tricky to follow.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6699)
Base 2^64 addition/subtraction and final reduction failed to treat
partially reduced values correctly.
Thanks to Wycheproof Project for vectors and Paul Kehrer for report.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6699)
Internal submodules of libcrypto may require non-public functions from
the EC submodule.
In preparation to use `ec_group_do_inverse_ord()` (from #6116) inside
the SM2 submodule to apply a SCA mitigation on the modular inversion,
this commit moves the `ec_group_do_inverse_ord()` prototype declaration
from the EC-local `crypto/ec/ec_lcl.h` header to the
`crypto/include/internal/ec_int.h` inter-module private header.
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6521)
BN_CTX_end() does not handle NULL input, so we must manually check
before calling from the cleanup handler.
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6502)
Fix prototype warnings triggered by -Wstrict-prototypes when configuring
with `enable-ec_nistp_64_gcc_128`
Reviewed-by: Kurt Roeckx <kurt@roeckx.be>
Reviewed-by: Matthias St. Pierre <Matthias.St.Pierre@ncp-e.com>
(Merged from https://github.com/openssl/openssl/pull/6556)
This extends the recently added ECDSA signature blinding to blind DSA too.
This is based on side channel attacks demonstrated by Keegan Ryan (NCC
Group) for ECDSA which are likely to be able to be applied to DSA.
Normally, as in ECDSA, during signing the signer calculates:
s:= k^-1 * (m + r * priv_key) mod order
In ECDSA, the addition operation above provides a sufficient signal for a
flush+reload attack to derive the private key given sufficient signature
operations.
As a mitigation (based on a suggestion from Keegan) we add blinding to
the operation so that:
s := k^-1 * blind^-1 (blind * m + blind * r * priv_key) mod order
Since this attack is a localhost side channel only no CVE is assigned.
This commit also tweaks the previous ECDSA blinding so that blinding is
only removed at the last possible step.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6522)
This commit implements coordinate blinding, i.e., it randomizes the
representative of an elliptic curve point in its equivalence class, for
prime curves implemented through EC_GFp_simple_method,
EC_GFp_mont_method, and EC_GFp_nist_method.
This commit is derived from the patch
https://marc.info/?l=openssl-dev&m=131194808413635 by Billy Brumley.
Coordinate blinding is a generally useful side-channel countermeasure
and is (mostly) free. The function itself takes a few field
multiplicationss, but is usually only necessary at the beginning of a
scalar multiplication (as implemented in the patch). When used this way,
it makes the values that variables take (i.e., field elements in an
algorithm state) unpredictable.
For instance, this mitigates chosen EC point side-channel attacks for
settings such as ECDH and EC private key decryption, for the
aforementioned curves.
For EC_METHODs using different coordinate representations this commit
does nothing, but the corresponding coordinate blinding function can be
easily added in the future to extend these changes to such curves.
Co-authored-by: Nicola Tuveri <nic.tuv@gmail.com>
Co-authored-by: Billy Brumley <bbrumley@gmail.com>
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6501)
Use EVP_PKEY_set_alias_type to access
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6443)
Keegan Ryan (NCC Group) has demonstrated a side channel attack on an
ECDSA signature operation. During signing the signer calculates:
s:= k^-1 * (m + r * priv_key) mod order
The addition operation above provides a sufficient signal for a
flush+reload attack to derive the private key given sufficient signature
operations.
As a mitigation (based on a suggestion from Keegan) we add blinding to
the operation so that:
s := k^-1 * blind^-1 (blind * m + blind * r * priv_key) mod order
Since this attack is a localhost side channel only no CVE is assigned.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Only applies to algorithms that support it. Both raw private and public
keys can be obtained for X25519, Ed25519, X448, Ed448. Raw private keys
only can be obtained for HMAC, Poly1305 and SipHash
Fixes#6259
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Tim Hudson <tjh@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6394)
Found by coverity. This is an artifact left over from the original
decaf import which generated the source code for different curves. For
curve 448 this is dead.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6373)
Return immediately upon discovery of bad message digest.
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6298)
This reverts commit a6f5b11634.
The EVP_PKEY_sign() function is intended for pre-hashed input which is
not supported by our EdDSA implementation.
See the discussion in PR 5880
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6284)
We check that the curve name associated with the point is the same as that
for the curve.
Fixes#6302
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6323)
Per SEC 1, the curve coefficients must be padded up to size. See C.2's
definition of Curve, C.1's definition of FieldElement, and 2.3.5's definition
of how to encode the field elements in http://www.secg.org/sec1-v2.pdf.
This comes up for P-521, where b needs a leading zero.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6314)
Using the ca application to sign certificates with EdDSA failed because it
is not possible to set the digest to "null". This adds the capability and
updates the documentation accordingly.
Fixes#6201
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6286)
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6070)
* EC_POINT_mul is now responsible for constant time point multiplication
(for single fixed or variable point multiplication, when the scalar is
in the range [0,group_order), so we need to strip the nonce padding
from ECDSA.
* Entry added to CHANGES
* Updated EC_POINT_mul documentation
- Integrate existing EC_POINT_mul and EC_POINTs_mul entries in the
manpage to reflect the shift in constant-time expectations when
performing a single fixed or variable point multiplication;
- Add documentation to ec_method_st to reflect the updated "contract"
between callers and implementations of ec_method_st.mul.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6070)
Co-authored-by: Nicola Tuveri <nic.tuv@gmail.com>
Co-authored-by: Cesar Pereida Garcia <cesar.pereidagarcia@tut.fi>
Co-authored-by: Sohaib ul Hassan <soh.19.hassan@gmail.com>
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6009)
Adding support for these operations for the EdDSA implementations
makes pkeyutl usable for signing/verifying for these algorithms.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/5880)
felem_neg does not produce an output within the tight bounds suitable
for felem_contract. This affects build configurations which set
enable-ec_nistp_64_gcc_128.
point_double and point_add, in the non-z*_is_zero cases, tolerate and
fix up the wider bounds, so this only affects point_add calls where the
other point is infinity. Thus it only affects the final addition in
arbitrary-point multiplication, giving the wrong y-coordinate. This is a
no-op for ECDH and ECDSA, which only use the x-coordinate of
arbitrary-point operations.
Note: ecp_nistp521.c has the same issue in that the documented
preconditions are violated by the test case. I have not addressed this
in this PR. ecp_nistp521.c does not immediately produce the wrong
answer; felem_contract there appears to be a bit more tolerant than its
documented preconditions. However, I haven't checked the point_add
property above holds. ecp_nistp521.c should either get this same fix, to
be conservative, or have the bounds analysis and comments reworked for
the wider bounds.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/5779)
Some platforms, cough-DJGPP, fail to compile claiming that requested
alignment is greater than maximum possible. Supposedly original
alignment was result of an attempt to utilize AVX2...
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/5708)
In particular, x and y may be NULL, as used in ecdsa_ossl.c. Make use of
this in ecdh_ossl.c as well, to save an otherwise unnecessary temporary.
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/5532)
Without actually using EVP_PKEY_FLAG_AUTOARGLEN
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4793)
Unlike "upstream", Android NDK's arm64 gcc [but not clang] performs
64x64=128-bit multiplications with library calls, which appears to
have devastating impact on performance. [The condition is reduced to
__ANDROID__ [&& !__clang__], because x86_64 has corresponding
assembly module.]
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/5589)
Debugging asserts had implicit casts that triggered the warnings.
However, instead of making the casts explicit it's more appropriate
to perform checks that ensure that implicit casts were safe.
ec/curve448/scalar.c: size_t-fy scalar_decode_short.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/5494)
This adds all of the relevant EVP plumbing required to make
X448 and Ed448 work.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Kurt Roeckx <kurt@roeckx.be>
(Merged from https://github.com/openssl/openssl/pull/5481)
Why is it redundant? We're looking at carry from addition of small,
11-bit number to 256-bit one. And carry would mean only one thing,
resulting first limb being small number and remaing ones - zeros.
Hence adding 38 to first limb can't carry.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/5476)
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/5449)
As it turns out gcc -pedantic doesn't seem to consider __uint128_t
as non-standard, unlike __int128 that is.
Fix even MSVC warnings in curve25519.c.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/5449)
SPARC condition in __SIZEOF_INT128__==16 is rather performance thing
than portability. Even though compiler advertises int128 capability,
corresponding operations are inefficient, because they are not
directly backed by instruction set.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/5449)
Currently it's limited to 64-bit platforms only as minimum radix
expected in assembly is 2^51.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/5408)
3 least significant bits of the input scalar are explicitly cleared,
hence swap variable has fixed value [of zero] upon exit from the loop.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/5408)
The original curve448 code was templated to allow for a 25519
implementation. We've just imported the 448 stuff - but a remnant of
the original templated approach remained. This just simplifies that.
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
(Merged from https://github.com/openssl/openssl/pull/5105)
