Timing vulnerability in DSA signature generation (CVE-2018-0734).

Avoid a timing attack that leaks information via a side channel that
triggers when a BN is resized.  Increasing the size of the BNs
prior to doing anything with them suppresses the attack.

Thanks due to Samuel Weiser for finding and locating this.

Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
(Merged from #7486)

(cherry picked from commit a9cfb8c)

crypto/dsa/dsa_ossl.c

@@ -9,6 +9,7 @@

#include <stdio.h>
#include "internal/cryptlib.h"
#include "internal/bn_int.h"
#include <openssl/bn.h>
#include <openssl/sha.h>
#include "dsa_locl.h"
@@ -180,9 +181,9 @@ static int dsa_sign_setup(DSA *dsa, BN_CTX *ctx_in,
{
    BN_CTX *ctx = NULL;
    BIGNUM *k, *kinv = NULL, *r = *rp;
    BIGNUM *l, *m;
    BIGNUM *l;
    int ret = 0;
    int q_bits;
    int q_bits, q_words;

    if (!dsa->p || !dsa->q || !dsa->g) {
        DSAerr(DSA_F_DSA_SIGN_SETUP, DSA_R_MISSING_PARAMETERS);
@@ -191,8 +192,7 @@ static int dsa_sign_setup(DSA *dsa, BN_CTX *ctx_in,

    k = BN_new();
    l = BN_new();
    m = BN_new();
    if (k == NULL || l == NULL || m == NULL)
    if (k == NULL || l == NULL)
        goto err;

    if (ctx_in == NULL) {
@@ -203,9 +203,9 @@ static int dsa_sign_setup(DSA *dsa, BN_CTX *ctx_in,

    /* Preallocate space */
    q_bits = BN_num_bits(dsa->q);
    if (!BN_set_bit(k, q_bits)
        || !BN_set_bit(l, q_bits)
        || !BN_set_bit(m, q_bits))
    q_words = bn_get_top(dsa->q);
    if (!bn_wexpand(k, q_words + 2)
        || !bn_wexpand(l, q_words + 2))
        goto err;

    /* Get random k */
@@ -240,14 +240,17 @@ static int dsa_sign_setup(DSA *dsa, BN_CTX *ctx_in,
     * small timing information leakage.  We then choose the sum that is
     * one bit longer than the modulus.
     *
     * TODO: revisit the BN_copy aiming for a memory access agnostic
     * conditional copy.
     * There are some concerns about the efficacy of doing this.  More
     * specificly refer to the discussion starting with:
     *     https://github.com/openssl/openssl/pull/7486#discussion_r228323705
     * The fix is to rework BN so these gymnastics aren't required.
     */
    if (!BN_add(l, k, dsa->q)
        || !BN_add(m, l, dsa->q)
        || !BN_copy(k, BN_num_bits(l) > q_bits ? l : m))
        || !BN_add(k, l, dsa->q))
        goto err;

    BN_consttime_swap(BN_is_bit_set(l, q_bits), k, l, q_words + 2);

    if ((dsa)->meth->bn_mod_exp != NULL) {
            if (!dsa->meth->bn_mod_exp(dsa, r, dsa->g, k, dsa->p, ctx,
                                       dsa->method_mont_p))
@@ -260,7 +263,7 @@ static int dsa_sign_setup(DSA *dsa, BN_CTX *ctx_in,
    if (!BN_mod(r, r, dsa->q, ctx))
        goto err;

    /* Compute  part of 's = inv(k) (m + xr) mod q' */
    /* Compute part of 's = inv(k) (m + xr) mod q' */
    if ((kinv = dsa_mod_inverse_fermat(k, dsa->q, ctx)) == NULL)
        goto err;

@@ -275,7 +278,6 @@ static int dsa_sign_setup(DSA *dsa, BN_CTX *ctx_in,
        BN_CTX_free(ctx);
    BN_clear_free(k);
    BN_clear_free(l);
    BN_clear_free(m);
    return ret;
}