We used to include platform.h only when MBEDTLS_PLATFORM_C was enabled, and
to define ad hoc replacements for mbedtls_xxx functions on a case-by-case
basis when MBEDTLS_PLATFORM_C was disabled. The only reason for this
complication was to allow building individual source modules without copying
platform.h. This is not something we support or recommend anymore, so get
rid of the complication: include platform.h unconditionally.
There should be no change in behavior since just including the header should
not change the behavior of a program.
This commit replaces most occurrences of conditional inclusion of
platform.h, using the following code:
```
perl -i -0777 -pe 's!#if.*\n#include "mbedtls/platform.h"\n(#else.*\n(#define (mbedtls|MBEDTLS)_.*\n|#include <(stdarg|stddef|stdio|stdlib|string|time)\.h>\n)*)?#endif.*!#include "mbedtls/platform.h"!mg' $(git grep -l '#include "mbedtls/platform.h"')
```
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
The bit length of m is leaked through through timing in ecp_mul_mxz().
Initially found by Manuel Pégourié-Gonnard on ecp_mul_edxyz(), which has
been inspired from ecp_mul_mxz(), during initial review of the EdDSA PR.
See: https://github.com/Mbed-TLS/mbedtls/pull/3245#discussion_r490827996
Fix that by using grp->nbits + 1 instead, which anyway is very close to
the length of m, which means there is no significant performance impact.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
These issues were flagged by Coverity as instances where a local
variable may be used prior to being initialized. Please note that
none of these changes fixes any particular bug, this is just an attempt
to add more robustness.
Signed-off-by: Leonid Rozenboim <leonid.rozenboim@oracle.com>
Previously, ecp_add_mixed(), commputing say P+Q, would allow for the
Q parameter to have an unset Z coordinate as a shortcut for Z == 1.
This was leveraged during computation and usage of the T-table
(storing low multiples of the to-be-multiplied point on the curve).
It is a potentially error-prone corner case, though, since an MPIs
with unset data pointer coordinate and limb size 0 is also a valid
representation of the number 0.
As a first step towards removing ECP points with unset Z coordinate,
the constant time T-array getter ecp_select_comb() has previously
been modified to return 'full' mbedtls_ecp_point structures,
including a 1-initialized Z-coordinate.
Similarly, this commit ...
- Modifies ecp_normalize_jac_many() to set the Z coordinates
of the points it operates on to 1 instead of freeing them.
- Frees the Z-coordinates of the T[]-array explicitly
once the computation and normalization of the T-table has finished.
As a minimal functional difference between old and new code,
the new code also frees the Z-coordinate of T[0]=P, which the
old code did not.
- Modifies ecp_add_mixed() to no longer allow unset Z coordinates.
Except for the post-precomputation storage form of the T[] array,
the code does therefore no longer use EC points with unset Z coordinate.
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
ecp_select_comb() did previously not set the Z coordinate of the target point.
Instead, callers would either set it explicitly or leave it uninitialized,
relying on the (only partly upheld) convention that sometimes an uninitialized
Z value represents 1.
This commit modifies ecp_select_comb() to always set the Z coordinate to 1.
This comes at the cost of memory for a single coordinate, which seems worth
it for the increased robustness.
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
This improves readibility and prepares for further changes
like the introduction of a single double-width temporary for
ECP arithmetic.
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
`ecp_add_mixed()` and `ecp_double_jac()` are the core subroutines
for elliptic curve arithmetic, and as such crucial for the performance
of ECP primitives like ECDHE and ECDSA.
This commit provides a very slight simplification and performance and
memory usage improvement to `ecp_add_mixed()` by removing the use of
three temporary MPIs used for coordinate calculations.
Where those variables were used, the code now writes directly to the
coordinate MPIs of the target elliptic curve point.
This is a valid change even if there is aliasing between input and
output, since at the time any of the coordinate MPIs in question is
written, the corresponding coordinates of both inputs are no longer
read.
(The analogous change in `ecp_double_jac()` can not be made since
this property does not hold for `ecp_double_jac()`.)
Signed-off-by: Hanno Becker <hanno.becker@arm.com>
minor changes, such as improving the documentation for the byte reading
macros, and using MBEDTLS_PUT_UINT16_xy in place of byte reading
macro combinations
Signed-off-by: Joe Subbiani <joe.subbiani@arm.com>
byte shifting opertations throughout library/ were only replaced with
the byte reading macros when an 0xff mask was being used.
The byte reading macros are now more widley used, however they have not
been used in all cases of a byte shift operation, as it detracted from
the immediate readability or otherwise did not seem appropriate.
Signed-off-by: Joe Subbiani <joe.subbiani@arm.com>
mbedtls_ecp_read_key and mbedtls_ecp_write_key are updated to include
support for Curve448 as prescribed by RFC 7748 §5.
Test suites have been updated to validate curve448 under Montgomery
curves.
Signed-off-by: Archana <archana.madhavan@silabs.com>
These macros were moved into a header and now check-names.sh is failing.
Add an MBEDTL_ prefix to the macro names to make it pass.
Signed-off-by: Janos Follath <janos.follath@arm.com>
We were already rejecting them at the end, due to the fact that with the
usual (x, z) formulas they lead to the result (0, 0) so when we want to
normalize at the end, trying to compute the modular inverse of z will
give an error.
If we wanted to support those points, we'd a special case in
ecp_normalize_mxz(). But it's actually permitted by all sources (RFC
7748 say we MAY reject 0 as a result) and recommended by some to reject
those points (either to ensure contributory behaviour, or to protect
against timing attack when the underlying field arithmetic is not
constant-time).
Since our field arithmetic is indeed not constant-time, let's reject
those points before they get mixed with sensitive data (in
ecp_mul_mxz()), in order to avoid exploitable leaks caused by the
special cases they would trigger. (See the "May the Fourth" paper
https://eprint.iacr.org/2017/806.pdf)
Signed-off-by: Janos Follath <janos.follath@arm.com>
Clang was complaining and check-names.sh too
This only duplicates macros, so no impact on code size. In 3.0 we can
probably avoid the duplication by using an internal header under
library/ but this won't work for 2.16.
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
We were already rejecting them at the end, due to the fact that with the
usual (x, z) formulas they lead to the result (0, 0) so when we want to
normalize at the end, trying to compute the modular inverse of z will
give an error.
If we wanted to support those points, we'd a special case in
ecp_normalize_mxz(). But it's actually permitted by all sources
(RFC 7748 say we MAY reject 0 as a result) and recommended by some to
reject those points (either to ensure contributory behaviour, or to
protect against timing attack when the underlying field arithmetic is
not constant-time).
Since our field arithmetic is indeed not constant-time, let's reject
those points before they get mixed with sensitive data (in
ecp_mul_mxz()), in order to avoid exploitable leaks caused by the
special cases they would trigger. (See the "May the Fourth" paper
https://eprint.iacr.org/2017/806.pdf)
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
