From db09fdc1a675bde167001a4f68e0f1e595e96dee Mon Sep 17 00:00:00 2001 From: Andy Polyakov Date: Fri, 10 Aug 2018 19:46:03 +0200 Subject: [PATCH] rsa/rsa_ossl.c: implement variant of "Smooth CRT-RSA." In [most common] case of p and q being of same width, it's possible to replace CRT modulo operations with Montgomery reductions. And those are even fixed-length Montgomery reductions... (cherry picked from commit 41bfd5e7c8ac3a0874a94e4d15c006ad5eb48e59) Resolved conflicts: crypto/rsa/rsa_ossl.c Reviewed-by: Paul Dale (Merged from https://github.com/openssl/openssl/pull/6940) --- crypto/rsa/rsa_ossl.c | 130 ++++++++++++++++++++++++++++++------------ 1 file changed, 94 insertions(+), 36 deletions(-) diff --git a/crypto/rsa/rsa_ossl.c b/crypto/rsa/rsa_ossl.c index 36c4e42a0f..5703411a89 100644 --- a/crypto/rsa/rsa_ossl.c +++ b/crypto/rsa/rsa_ossl.c @@ -127,8 +127,8 @@ static int rsa_ossl_public_encrypt(int flen, const unsigned char *from, } if (rsa->flags & RSA_FLAG_CACHE_PUBLIC) - if (!BN_MONT_CTX_set_locked - (&rsa->_method_mod_n, rsa->lock, rsa->n, ctx)) + if (!BN_MONT_CTX_set_locked(&rsa->_method_mod_n, rsa->lock, + rsa->n, ctx)) goto err; if (!rsa->meth->bn_mod_exp(ret, f, rsa->e, rsa->n, ctx, @@ -312,8 +312,8 @@ static int rsa_ossl_private_encrypt(int flen, const unsigned char *from, BN_with_flags(d, rsa->d, BN_FLG_CONSTTIME); if (rsa->flags & RSA_FLAG_CACHE_PUBLIC) - if (!BN_MONT_CTX_set_locked - (&rsa->_method_mod_n, rsa->lock, rsa->n, ctx)) { + if (!BN_MONT_CTX_set_locked(&rsa->_method_mod_n, rsa->lock, + rsa->n, ctx)) { BN_free(d); goto err; } @@ -435,8 +435,8 @@ static int rsa_ossl_private_decrypt(int flen, const unsigned char *from, BN_with_flags(d, rsa->d, BN_FLG_CONSTTIME); if (rsa->flags & RSA_FLAG_CACHE_PUBLIC) - if (!BN_MONT_CTX_set_locked - (&rsa->_method_mod_n, rsa->lock, rsa->n, ctx)) { + if (!BN_MONT_CTX_set_locked(&rsa->_method_mod_n, rsa->lock, + rsa->n, ctx)) { BN_free(d); goto err; } @@ -541,8 +541,8 @@ static int rsa_ossl_public_decrypt(int flen, const unsigned char *from, } if (rsa->flags & RSA_FLAG_CACHE_PUBLIC) - if (!BN_MONT_CTX_set_locked - (&rsa->_method_mod_n, rsa->lock, rsa->n, ctx)) + if (!BN_MONT_CTX_set_locked(&rsa->_method_mod_n, rsa->lock, + rsa->n, ctx)) goto err; if (!rsa->meth->bn_mod_exp(ret, f, rsa->e, rsa->n, ctx, @@ -583,7 +583,7 @@ static int rsa_ossl_public_decrypt(int flen, const unsigned char *from, static int rsa_ossl_mod_exp(BIGNUM *r0, const BIGNUM *I, RSA *rsa, BN_CTX *ctx) { BIGNUM *r1, *m1, *vrfy; - int ret = 0; + int ret = 0, smooth = 0; BN_CTX_start(ctx); @@ -593,43 +593,79 @@ static int rsa_ossl_mod_exp(BIGNUM *r0, const BIGNUM *I, RSA *rsa, BN_CTX *ctx) if (vrfy == NULL) goto err; - { - BIGNUM *p = BN_new(), *q = BN_new(); + 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 (p == NULL || q == NULL) { - BN_free(p); - BN_free(q); + 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; } - BN_with_flags(p, rsa->p, BN_FLG_CONSTTIME); - BN_with_flags(q, rsa->q, BN_FLG_CONSTTIME); - - if (rsa->flags & RSA_FLAG_CACHE_PRIVATE) { - if (!BN_MONT_CTX_set_locked - (&rsa->_method_mod_p, rsa->lock, p, ctx) - || !BN_MONT_CTX_set_locked(&rsa->_method_mod_q, - rsa->lock, q, ctx)) { - BN_free(p); - BN_free(q); - goto err; - } - } /* - * We MUST free p and q before any further use of rsa->p and rsa->q + * We MUST free |factor| before any further use of the prime factors */ - BN_free(p); - BN_free(q); + BN_free(factor); + + smooth = (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)) + 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) + + /* r0 = r0 * 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) + || !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(); @@ -652,7 +688,7 @@ static int rsa_ossl_mod_exp(BIGNUM *r0, const BIGNUM *I, RSA *rsa, BN_CTX *ctx) /* compute r1^dmq1 mod q */ if (!rsa->meth->bn_mod_exp(m1, r1, dmq1, rsa->q, ctx, - rsa->_method_mod_q)) { + rsa->_method_mod_q)) { BN_free(c); BN_free(dmq1); goto err; @@ -728,10 +764,18 @@ static int rsa_ossl_mod_exp(BIGNUM *r0, const BIGNUM *I, RSA *rsa, BN_CTX *ctx) if (!BN_add(r0, r1, m1)) goto err; + tail: if (rsa->e && rsa->n) { - if (!rsa->meth->bn_mod_exp(vrfy, r0, rsa->e, rsa->n, ctx, - rsa->_method_mod_n)) - goto err; + 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 @@ -740,6 +784,11 @@ static int rsa_ossl_mod_exp(BIGNUM *r0, const BIGNUM *I, RSA *rsa, BN_CTX *ctx) */ 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)) @@ -766,6 +815,15 @@ static int rsa_ossl_mod_exp(BIGNUM *r0, const BIGNUM *I, RSA *rsa, BN_CTX *ctx) 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); -- 2.25.1