int BN_mod_mul_montgomery(BIGNUM *r, const BIGNUM *a, const BIGNUM *b,
BN_MONT_CTX *mont, BN_CTX *ctx)
{
int ret = bn_mul_mont_fixed_top(r, a, b, mont, ctx);
bn_correct_top(r);
bn_check_top(r);
return ret;
}
static ECDSA_SIG *ecdsa_do_sign(const unsigned char *dgst, int dgst_len,
const BIGNUM *in_kinv, const BIGNUM *in_r,
EC_KEY *eckey)
{
int ok = 0, i;
BIGNUM *kinv = NULL, *s, *m = NULL, *order = NULL;
const BIGNUM *ckinv;
BN_CTX *ctx = NULL;
const EC_GROUP *group;
ECDSA_SIG *ret;
ECDSA_DATA *ecdsa;
const BIGNUM *priv_key;
BN_MONT_CTX *mont_data;
ecdsa = ecdsa_check(eckey);
group = EC_KEY_get0_group(eckey);
priv_key = EC_KEY_get0_private_key(eckey);
if (group == NULL || priv_key == NULL || ecdsa == NULL) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN, ERR_R_PASSED_NULL_PARAMETER);
return NULL;
}
ret = ECDSA_SIG_new();
if (!ret) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN, ERR_R_MALLOC_FAILURE);
return NULL;
}
s = ret->s;
if ((ctx = BN_CTX_new()) == NULL || (order = BN_new()) == NULL ||
(m = BN_new()) == NULL) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN, ERR_R_MALLOC_FAILURE);
goto err;
}
if (!EC_GROUP_get_order(group, order, ctx)) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN, ERR_R_EC_LIB);
goto err;
}
mont_data = EC_GROUP_get_mont_data(group);
i = BN_num_bits(order);
/*
* Need to truncate digest if it is too long: first truncate whole bytes.
*/
if (8 * dgst_len > i)
dgst_len = (i + 7) / 8;
if (!BN_bin2bn(dgst, dgst_len, m)) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN, ERR_R_BN_LIB);
goto err;
}
/* If still too long truncate remaining bits with a shift */
if ((8 * dgst_len > i) && !BN_rshift(m, m, 8 - (i & 0x7))) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN, ERR_R_BN_LIB);
goto err;
}
do {
if (in_kinv == NULL || in_r == NULL) {
if (!ECDSA_sign_setup(eckey, ctx, &kinv, &ret->r)) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN, ERR_R_ECDSA_LIB);
goto err;
}
ckinv = kinv;
} else {
ckinv = in_kinv;
if (BN_copy(ret->r, in_r) == NULL) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN, ERR_R_MALLOC_FAILURE);
goto err;
}
}
/*
* With only one multiplicant being in Montgomery domain
* multiplication yields real result without post-conversion.
* Also note that all operations but last are performed with
* zero-padded vectors. Last operation, BN_mod_mul_montgomery
* below, returns user-visible value with removed zero padding.
*/
if (!bn_to_mont_fixed_top(s, ret->r, mont_data, ctx)
|| !bn_mul_mont_fixed_top(s, s, priv_key, mont_data, ctx)) {
goto err;
}
if (!bn_mod_add_fixed_top(s, s, m, order)) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN, ERR_R_BN_LIB);
goto err;
}
/*
* |s| can still be larger than modulus, because |m| can be. In
* such case we count on Montgomery reduction to tie it up.
*/
if (!bn_to_mont_fixed_top(s, s, mont_data, ctx)
|| !BN_mod_mul_montgomery(s, s, ckinv, mont_data, ctx)) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN, ERR_R_BN_LIB);
goto err;
}
if (BN_is_zero(s)) {
/*
* if kinv and r have been supplied by the caller don't to
* generate new kinv and r values
*/
if (in_kinv != NULL && in_r != NULL) {
ECDSAerr(ECDSA_F_ECDSA_DO_SIGN,
ECDSA_R_NEED_NEW_SETUP_VALUES);
goto err;
}
} else
/* s != 0 => we have a valid signature */
break;
}
while (1);
ok = 1;
err:
if (!ok) {
ECDSA_SIG_free(ret);
ret = NULL;
}
if (ctx)
BN_CTX_free(ctx);
if (m)
BN_clear_free(m);
if (order)
BN_free(order);
if (kinv)
BN_clear_free(kinv);
return ret;
}
static int rsa_ossl_mod_exp(BIGNUM *r0, const BIGNUM *I, RSA *rsa, BN_CTX *ctx)
{
BIGNUM *r1, *m1, *vrfy, *r2, *m[RSA_MAX_PRIME_NUM - 2];
int ret = 0, i, ex_primes = 0, smooth = 0;
RSA_PRIME_INFO *pinfo;
BN_CTX_start(ctx);
r1 = BN_CTX_get(ctx);
r2 = BN_CTX_get(ctx);
m1 = BN_CTX_get(ctx);
vrfy = BN_CTX_get(ctx);
if (vrfy == NULL)
goto err;
if (rsa->version == RSA_ASN1_VERSION_MULTI
&& ((ex_primes = sk_RSA_PRIME_INFO_num(rsa->prime_infos)) <= 0
|| ex_primes > RSA_MAX_PRIME_NUM - 2))
goto err;
if (rsa->flags & RSA_FLAG_CACHE_PRIVATE) {
BIGNUM *factor = BN_new();
if (factor == NULL)
goto err;
/*
* Make sure BN_mod_inverse in Montgomery initialization uses the
* BN_FLG_CONSTTIME flag
*/
if (!(BN_with_flags(factor, rsa->p, BN_FLG_CONSTTIME),
BN_MONT_CTX_set_locked(&rsa->_method_mod_p, rsa->lock,
factor, ctx))
|| !(BN_with_flags(factor, rsa->q, BN_FLG_CONSTTIME),
BN_MONT_CTX_set_locked(&rsa->_method_mod_q, rsa->lock,
factor, ctx))) {
BN_free(factor);
goto err;
}
for (i = 0; i < ex_primes; i++) {
pinfo = sk_RSA_PRIME_INFO_value(rsa->prime_infos, i);
BN_with_flags(factor, pinfo->r, BN_FLG_CONSTTIME);
if (!BN_MONT_CTX_set_locked(&pinfo->m, rsa->lock, factor, ctx)) {
BN_free(factor);
goto err;
}
}
/*
* We MUST free |factor| before any further use of the prime factors
*/
BN_free(factor);
smooth = (ex_primes == 0)
&& (rsa->meth->bn_mod_exp == BN_mod_exp_mont)
&& (BN_num_bits(rsa->q) == BN_num_bits(rsa->p));
}
if (rsa->flags & RSA_FLAG_CACHE_PUBLIC)
if (!BN_MONT_CTX_set_locked(&rsa->_method_mod_n, rsa->lock,
rsa->n, ctx))
goto err;
if (smooth) {
/*
* Conversion from Montgomery domain, a.k.a. Montgomery reduction,
* accepts values in [0-m*2^w) range. w is m's bit width rounded up
* to limb width. So that at the very least if |I| is fully reduced,
* i.e. less than p*q, we can count on from-to round to perform
* below modulo operations on |I|. Unlike BN_mod it's constant time.
*/
if (/* m1 = I moq q */
!bn_from_mont_fixed_top(m1, I, rsa->_method_mod_q, ctx)
|| !bn_to_mont_fixed_top(m1, m1, rsa->_method_mod_q, ctx)
/* m1 = m1^dmq1 mod q */
|| !BN_mod_exp_mont_consttime(m1, m1, rsa->dmq1, rsa->q, ctx,
rsa->_method_mod_q)
/* r1 = I mod p */
|| !bn_from_mont_fixed_top(r1, I, rsa->_method_mod_p, ctx)
|| !bn_to_mont_fixed_top(r1, r1, rsa->_method_mod_p, ctx)
/* r1 = r1^dmp1 mod p */
|| !BN_mod_exp_mont_consttime(r1, r1, rsa->dmp1, rsa->p, ctx,
rsa->_method_mod_p)
/* r1 = (r1 - m1) mod p */
/*
* bn_mod_sub_fixed_top is not regular modular subtraction,
* it can tolerate subtrahend to be larger than modulus, but
* not bit-wise wider. This makes up for uncommon q>p case,
* when |m1| can be larger than |rsa->p|.
*/
|| !bn_mod_sub_fixed_top(r1, r1, m1, rsa->p)
/* r1 = r1 * iqmp mod p */
|| !bn_to_mont_fixed_top(r1, r1, rsa->_method_mod_p, ctx)
|| !bn_mul_mont_fixed_top(r1, r1, rsa->iqmp, rsa->_method_mod_p,
ctx)
/* r0 = r1 * q + m1 */
|| !bn_mul_fixed_top(r0, r1, rsa->q, ctx)
|| !bn_mod_add_fixed_top(r0, r0, m1, rsa->n))
goto err;
goto tail;
}
/* compute I mod q */
{
BIGNUM *c = BN_new();
if (c == NULL)
goto err;
BN_with_flags(c, I, BN_FLG_CONSTTIME);
if (!BN_mod(r1, c, rsa->q, ctx)) {
BN_free(c);
goto err;
}
{
BIGNUM *dmq1 = BN_new();
if (dmq1 == NULL) {
BN_free(c);
goto err;
}
BN_with_flags(dmq1, rsa->dmq1, BN_FLG_CONSTTIME);
/* compute r1^dmq1 mod q */
if (!rsa->meth->bn_mod_exp(m1, r1, dmq1, rsa->q, ctx,
rsa->_method_mod_q)) {
BN_free(c);
BN_free(dmq1);
goto err;
}
/* We MUST free dmq1 before any further use of rsa->dmq1 */
BN_free(dmq1);
}
/* compute I mod p */
if (!BN_mod(r1, c, rsa->p, ctx)) {
BN_free(c);
goto err;
}
/* We MUST free c before any further use of I */
BN_free(c);
}
{
BIGNUM *dmp1 = BN_new();
if (dmp1 == NULL)
goto err;
BN_with_flags(dmp1, rsa->dmp1, BN_FLG_CONSTTIME);
/* compute r1^dmp1 mod p */
if (!rsa->meth->bn_mod_exp(r0, r1, dmp1, rsa->p, ctx,
rsa->_method_mod_p)) {
BN_free(dmp1);
goto err;
}
/* We MUST free dmp1 before any further use of rsa->dmp1 */
BN_free(dmp1);
}
/*
* calculate m_i in multi-prime case
*
* TODO:
* 1. squash the following two loops and calculate |m_i| there.
* 2. remove cc and reuse |c|.
* 3. remove |dmq1| and |dmp1| in previous block and use |di|.
*
* If these things are done, the code will be more readable.
*/
if (ex_primes > 0) {
BIGNUM *di = BN_new(), *cc = BN_new();
if (cc == NULL || di == NULL) {
BN_free(cc);
BN_free(di);
goto err;
}
for (i = 0; i < ex_primes; i++) {
/* prepare m_i */
if ((m[i] = BN_CTX_get(ctx)) == NULL) {
BN_free(cc);
BN_free(di);
goto err;
}
pinfo = sk_RSA_PRIME_INFO_value(rsa->prime_infos, i);
/* prepare c and d_i */
BN_with_flags(cc, I, BN_FLG_CONSTTIME);
BN_with_flags(di, pinfo->d, BN_FLG_CONSTTIME);
if (!BN_mod(r1, cc, pinfo->r, ctx)) {
BN_free(cc);
BN_free(di);
goto err;
}
/* compute r1 ^ d_i mod r_i */
if (!rsa->meth->bn_mod_exp(m[i], r1, di, pinfo->r, ctx, pinfo->m)) {
BN_free(cc);
BN_free(di);
goto err;
}
}
BN_free(cc);
BN_free(di);
}
if (!BN_sub(r0, r0, m1))
goto err;
/*
* This will help stop the size of r0 increasing, which does affect the
* multiply if it optimised for a power of 2 size
*/
if (BN_is_negative(r0))
if (!BN_add(r0, r0, rsa->p))
goto err;
if (!BN_mul(r1, r0, rsa->iqmp, ctx))
goto err;
{
BIGNUM *pr1 = BN_new();
if (pr1 == NULL)
goto err;
BN_with_flags(pr1, r1, BN_FLG_CONSTTIME);
if (!BN_mod(r0, pr1, rsa->p, ctx)) {
BN_free(pr1);
goto err;
}
/* We MUST free pr1 before any further use of r1 */
BN_free(pr1);
}
/*
* If p < q it is occasionally possible for the correction of adding 'p'
* if r0 is negative above to leave the result still negative. This can
* break the private key operations: the following second correction
* should *always* correct this rare occurrence. This will *never* happen
* with OpenSSL generated keys because they ensure p > q [steve]
*/
if (BN_is_negative(r0))
if (!BN_add(r0, r0, rsa->p))
goto err;
if (!BN_mul(r1, r0, rsa->q, ctx))
goto err;
if (!BN_add(r0, r1, m1))
goto err;
/* add m_i to m in multi-prime case */
if (ex_primes > 0) {
BIGNUM *pr2 = BN_new();
if (pr2 == NULL)
goto err;
for (i = 0; i < ex_primes; i++) {
pinfo = sk_RSA_PRIME_INFO_value(rsa->prime_infos, i);
if (!BN_sub(r1, m[i], r0)) {
BN_free(pr2);
goto err;
}
if (!BN_mul(r2, r1, pinfo->t, ctx)) {
BN_free(pr2);
goto err;
}
BN_with_flags(pr2, r2, BN_FLG_CONSTTIME);
if (!BN_mod(r1, pr2, pinfo->r, ctx)) {
BN_free(pr2);
goto err;
}
if (BN_is_negative(r1))
if (!BN_add(r1, r1, pinfo->r)) {
BN_free(pr2);
goto err;
}
if (!BN_mul(r1, r1, pinfo->pp, ctx)) {
BN_free(pr2);
goto err;
}
if (!BN_add(r0, r0, r1)) {
BN_free(pr2);
goto err;
}
}
BN_free(pr2);
}
tail:
if (rsa->e && rsa->n) {
if (rsa->meth->bn_mod_exp == BN_mod_exp_mont) {
if (!BN_mod_exp_mont(vrfy, r0, rsa->e, rsa->n, ctx,
rsa->_method_mod_n))
goto err;
} else {
bn_correct_top(r0);
if (!rsa->meth->bn_mod_exp(vrfy, r0, rsa->e, rsa->n, ctx,
rsa->_method_mod_n))
goto err;
}
/*
* If 'I' was greater than (or equal to) rsa->n, the operation will
* be equivalent to using 'I mod n'. However, the result of the
* verify will *always* be less than 'n' so we don't check for
* absolute equality, just congruency.
*/
if (!BN_sub(vrfy, vrfy, I))
goto err;
if (BN_is_zero(vrfy)) {
bn_correct_top(r0);
ret = 1;
goto err; /* not actually error */
}
if (!BN_mod(vrfy, vrfy, rsa->n, ctx))
goto err;
if (BN_is_negative(vrfy))
if (!BN_add(vrfy, vrfy, rsa->n))
goto err;
if (!BN_is_zero(vrfy)) {
/*
* 'I' and 'vrfy' aren't congruent mod n. Don't leak
* miscalculated CRT output, just do a raw (slower) mod_exp and
* return that instead.
*/
BIGNUM *d = BN_new();
if (d == NULL)
goto err;
BN_with_flags(d, rsa->d, BN_FLG_CONSTTIME);
if (!rsa->meth->bn_mod_exp(r0, I, d, rsa->n, ctx,
rsa->_method_mod_n)) {
BN_free(d);
goto err;
}
/* We MUST free d before any further use of rsa->d */
BN_free(d);
}
}
/*
* It's unfortunate that we have to bn_correct_top(r0). What hopefully
* saves the day is that correction is highly unlike, and private key
* operations are customarily performed on blinded message. Which means
* that attacker won't observe correlation with chosen plaintext.
* Secondly, remaining code would still handle it in same computational
* time and even conceal memory access pattern around corrected top.
*/
bn_correct_top(r0);
ret = 1;
err:
BN_CTX_end(ctx);
return ret;
}
int bn_to_mont_fixed_top(BIGNUM *r, const BIGNUM *a, BN_MONT_CTX *mont,
BN_CTX *ctx)
{
return bn_mul_mont_fixed_top(r, a, &(mont->RR), mont, ctx);
}