The input distribution to primality testing functions is completely
different when used for generating primes and when for validating
primes. The constants used in the library are geared towards the prime
generation use case and are weak when used for validation. (Maliciously
constructed composite numbers can pass the test with high probability)
The mbedtls_mpi_is_prime() function is in the public API and although it
is not documented, it is reasonable to assume that the primary use case
is validating primes. The RSA module too uses it for validating key
material.
Primality tests have to deal with different distribution when generating
primes and when validating primes.
These new tests are testing if mbedtls_mpi_is_prime() is working
properly in the latter setting.
The new tests involve pseudoprimes with maximum number of
non-witnesses. The non-witnesses were generated by printing them
from mpi_miller_rabin(). The pseudoprimes were generated by the
following function:
void gen_monier( mbedtls_mpi* res, int nbits )
{
mbedtls_mpi p_2x_plus_1, p_4x_plus_1, x, tmp;
mbedtls_mpi_init( &p_2x_plus_1 );
mbedtls_mpi_init( &p_4x_plus_1 );
mbedtls_mpi_init( &x ); mbedtls_mpi_init( &tmp );
do
{
mbedtls_mpi_gen_prime( &p_2x_plus_1, nbits >> 1, 0,
rnd_std_rand, NULL );
mbedtls_mpi_sub_int( &x, &p_2x_plus_1, 1 );
mbedtls_mpi_div_int( &x, &tmp, &x, 2 );
if( mbedtls_mpi_get_bit( &x, 0 ) == 0 )
continue;
mbedtls_mpi_mul_int( &p_4x_plus_1, &x, 4 );
mbedtls_mpi_add_int( &p_4x_plus_1, &p_4x_plus_1, 1 );
if( mbedtls_mpi_is_prime( &p_4x_plus_1, rnd_std_rand,
NULL ) == 0 )
break;
} while( 1 );
mbedtls_mpi_mul_mpi( res, &p_2x_plus_1, &p_4x_plus_1 );
}
The FIPS 186-4 RSA key generation prescribes lower failure probability
in primality testing and this makes key generation slower. We enable the
caller to decide between compliance/security and performance.
This python script calculates the base two logarithm of the formulas in
HAC Fact 4.48 and was used to determine the breakpoints and number of
rounds:
def mrpkt_log_2(k, t):
if t <= k/9.0:
return 3*math.log(k,2)/2+t-math.log(t,2)/2+4-2*math.sqrt(t*k)
elif t <= k/4.0:
c1 = math.log(7.0*k/20,2)-5*t
c2 = math.log(1/7.0,2)+15*math.log(k,2)/4.0-k/2.0-2*t
c3 = math.log(12*k,2)-k/4.0-3*t
return max(c1, c2, c3)
else:
return math.log(1/7.0)+15*math.log(k,2)/4.0-k/2.0-2*t
If `MBEDTLS_MEMORY_BUFFER_ALLOC_C` is configured and Mbed TLS'
custom buffer allocator is used for calloc() and free(), the
read buffer used by the server example application is allocated
from the buffer allocator, but freed after the buffer allocator
has been destroyed. If memory backtracing is enabled, this leaves
a memory leak in the backtracing structure allocated for the buffer,
as found by valgrind.
Fixes#2069.
In the previous bounds check `(*p) > end - len`, the computation
of `end - len` might underflow if `end` is within the first 64KB
of the address space (note that the length `len` is controlled by
the peer). In this case, the bounds check will be bypassed, leading
to `*p` exceed the message bounds by up to 64KB when leaving
`ssl_parse_server_psk_hint()`. In a pure PSK-based handshake,
this doesn't seem to have any consequences, as `*p*` is not accessed
afterwards. In a PSK-(EC)DHE handshake, however, `*p` is read from
in `ssl_parse_server_ecdh_params()` and `ssl_parse_server_dh_params()`
which might lead to an application crash of information leakage.
Functional tests for various payload sizes and output buffer sizes.
When the padding is bad or the plaintext is too large for the output
buffer, verify that function writes some outputs. This doesn't
validate that the implementation is time-constant, but it at least
validates that it doesn't just return early without outputting anything.
Get rid of the variable p. This makes it more apparent where the code
accesses the buffer at an offset whose value is sensitive.
No intended behavior change in this commit.
* The variables `csr` and `issuer_crt` are initialized but not freed.
* The variable `entropy` is unconditionally freed in the cleanup section
but there's a conditional jump to that section before its initialization.
This cmmot Moves it to the other initializations happening before the
first conditional jump to the cleanup section.
Fixes#1422.
Rather than doing the quadratic-time constant-memory-trace on the
whole working buffer, do it on the section of the buffer where the
data to copy has to lie, which can be significantly smaller if the
output buffer is significantly smaller than the working buffer, e.g.
for TLS RSA ciphersuites (48 bytes vs MBEDTLS_MPI_MAX_SIZE).
In mbedtls_rsa_rsaes_pkcs1_v15_decrypt, use size_greater_than (which
is based on bitwise operations) instead of the < operator to compare
sizes when the values being compared must not leak. Some compilers
compile < to a branch at least under some circumstances (observed with
gcc 5.4 for arm-gnueabi -O9 on a toy program).
Replace memmove(to, to + offset, length) by a functionally equivalent
function that strives to make the same memory access patterns
regardless of the value of length. This fixes an information leak
through timing (especially timing of memory accesses via cache probes)
that leads to a Bleichenbacher-style attack on PKCS#1 v1.5 decryption
using the plaintext length as the observable.
mbedtls_rsa_rsaes_pkcs1_v15_decrypt takes care not to reveal whether
the padding is valid or not, even through timing or memory access
patterns. This is a defense against an attack published by
Bleichenbacher. The attacker can also obtain the same information by
observing the length of the plaintext. The current implementation
leaks the length of the plaintext through timing and memory access
patterns.
This commit is a first step towards fixing this leak. It reduces the
leak to a single memmove call inside the working buffer.
Make the function more robust by taking an arbitrary zero/nonzero
argument instead of insisting on zero/all-bits-one. Update and fix its
documentation.
The Cortex M4, M7 MCUs and the Cortex A CPUs support the ARM DSP
instructions, and especially the umaal instruction which greatly
speed up MULADDC code. In addition the patch switched the ASM
constraints to registers instead of memory, giving the opportunity
for the compiler to load them the best way.
The speed improvement is variable depending on the crypto operation
and the CPU. Here are the results on a Cortex M4, a Cortex M7 and a
Cortex A8. All tests have been done with GCC 6.3 using -O2. RSA uses a
RSA-4096 key. ECDSA uses a secp256r1 curve EC key pair.
+--------+--------+--------+
| M4 | M7 | A8 |
+----------------+--------+--------+--------+
| ECDSA signing | +6.3% | +7.9% | +4.1% |
+----------------+--------+--------+--------+
| RSA signing | +43.7% | +68.3% | +26.3% |
+----------------+--------+--------+--------+
| RSA encryption | +3.4% | +9.7% | +3.6% |
+----------------+--------+--------+--------+
| RSA decryption | +43.0% | +67.8% | +22.8% |
+----------------+--------+--------+--------+
I ran the whole testsuite on the Cortex A8 Linux environment, and it
all passes.
stdio.h was being included both conditionally if MBEDTLS_FS_IO was
defined, and also unconditionally, which made at least one of them
redundant.
This change removes the unconditional inclusion of stdio.h and makes it
conditional on MBEDTLS_PLATFORM_C.
mbedtls_rsa_rsaes_pkcs1_v15_decrypt took care of calculating the
padding length without leaking the amount of padding or the validity
of the padding. However it then skipped the copying of the data if the
padding was invalid, which could allow an adversary to find out
whether the padding was valid through precise timing measurements,
especially if for a local attacker who could observe memory access via
cache timings.
Avoid this leak by always copying from the decryption buffer to the
output buffer, even when the padding is invalid. With invalid padding,
copy the same amount of data as what is expected on valid padding: the
minimum valid padding size if this fits in the output buffer,
otherwise the output buffer size. To avoid leaking payload data from
an unsuccessful decryption, zero the decryption buffer before copying
if the padding was invalid.
Exclude ".git" directories anywhere. This avoids spurious errors in git
checkouts that contain branch names that look like a file
check-files.py would check. Fix#1713
Exclude "mbed-os" anywhere and "examples" from the root. Switch to the
new mechanism to exclude "yotta/module". These are directories where
we store third-party files that do not need to match our preferences.
Exclude "cov-int" from the root. Fix#1691
Changes run-test-suites.pl to filter out directories, and select only files
as on OSX, test coverage tests create .dSYM directories which were being
accidentally selected to execute.
Remove the trailing whitespace from the inline assembly for AMD64 target, to
overcome a warning in Clang, which was objecting to the string literal
generated by the inline assembly being greater than 4096 characters specified
by the ISO C99 standard. (-Woverlength-strings)
This is a cosmetic change and doesn't change the logic of the code in any way.
This change only fixes the problem for AMD64 target, and leaves other targets as
they are.
Fixes#482.
We don't need to disable ASLR, so don't try. If gdb tries but fails,
the test runs normally, but all.sh then trips up because it sees
`warning: Error disabling address space randomization: Operation not permitted`
and interprets it as an error that indicates a test failure.
Generate the documentation from include and doxygen/input only. Don't
get snared by files containing Doxygen comments that lie in other
directories such as tests, yotta, crypto/include, ...
The only difference this makes in a fresh checkout is that the
documentation no longer lists target_config.h. This file is from
yotta, does not contain any Doxygen comment, and its inclusion in the
rendered documentation was clearly an oversight.
Add signing tests with 528-bit and 520-bit RSA keys with SHA-512. These
selections of key and hash size should lead to an error returned, as
there is not enough room for our chosen minimum salt size of two bytes
less than the hash size. These test the boundary around an available
salt length of 0 or -1 bytes.
The RSA keys were generated with OpenSSL 1.1.1-pre8.
$ openssl genrsa 520
Generating RSA private key, 520 bit long modulus (2 primes)
.............++++++++++++
.................++++++++++++
e is 65537 (0x010001)
-----BEGIN RSA PRIVATE KEY-----
MIIBPwIBAAJCANWgb4bludh0KFQBZcqWb6iJOmLipZ0L/XYXeAuwOfkWWjc6jhGd
B2b43lVnEPM/ZwGRU7rYIjd155fUUdSCBvO/AgMBAAECQgDOMq+zy6XZEjWi8D5q
j05zpRGgRRiKP/qEtB6BWbZ7gUV9DDgZhD4FFsqfanwjWNG52LkM9D1OQmUOtGGq
a9COwQIhD+6l9iIPrCkblQjsK6jtKB6zmu5NXcaTJUEGgW68cA7PAiENaJGHhcOq
/jHqqi2NgVbc5kWUD/dzSkVzN6Ub0AvIiBECIQIeL2Gw1XSFYm1Fal/DbQNQUX/e
/dnhc94X7s118wbScQIhAMPVgbDc//VurZ+155vYc9PjZlYe3QIAwlkLX3HYKkGx
AiEND8ndKyhkc8jLGlh8aRP8r03zpDIiZNKqCKiijMWVRYQ=
-----END RSA PRIVATE KEY-----
$ openssl genrsa 528
Generating RSA private key, 528 bit long modulus (2 primes)
.........++++++++++++
....++++++++++++
e is 65537 (0x010001)
-----BEGIN RSA PRIVATE KEY-----
MIIBQgIBAAJDAKJVTrpxW/ZuXs3z1tcY4+XZB+hmbnv1p2tBUQbgTrgn7EyyGZz/
ZkkdRUGQggWapbVLDPXu9EQ0AvMEfAsObwJQgQIDAQABAkJhHVXvFjglElxnK7Rg
lERq0k73yqfYQts4wCegTHrrkv3HzqWQVVi29mGLSXTqoQ45gzWZ5Ru5NKjkTjko
YtWWIVECIgDScqoo7SCFrG3zwFxnGe7V3rYYr6LkykpvczC0MK1IZy0CIgDFeINr
qycUXbndZvF0cLYtSmEA+MoN7fRX7jY5w7lZYyUCIUxyiOurEDhe5eY5B5gQbJlW
ePHIw7S244lO3+9lC12U1QIhWgzQ8YKFObZcEejl5xGXIiQvBEBv89Y1fPu2YrUs
iuS5AiFE64NJs8iI+zZxp72esKHPXq/chJ1BvhHsXI0y1OBK8m8=
-----END RSA PRIVATE KEY-----
