Elliptic curve scalar multiplication with timing attack defenses

Co-authored-by: Nicola Tuveri <nic.tuv@gmail.com>
Co-authored-by: Cesar Pereida Garcia <cesar.pereidagarcia@tut.fi>
Co-authored-by: Sohaib ul Hassan <soh.19.hassan@gmail.com>
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6009)

diff --git a/crypto/bn/bn_lib.c b/crypto/bn/bn_lib.c
index 57fe45288bbabb3dfb565e42962ea3507d641a7a..a446880ec727228a3a44986b9cad73e3d4600b61 100644 (file)
--- a/crypto/bn/bn_lib.c
+++ b/crypto/bn/bn_lib.c
@@ -739,6 +739,19 @@ void BN_consttime_swap(BN_ULONG condition, BIGNUM *a, BIGNUM *b, int nwords)
     a->top ^= t;
     b->top ^= t;
 
+    t = (a->neg ^ b->neg) & condition;
+    a->neg ^= t;
+    b->neg ^= t;
+
+    /*
+     * cannot just arbitrarily swap flags.
+     * The way a->d is allocated etc.
+     * BN_FLG_MALLOCED, BN_FLG_STATIC_DATA, ...
+     */
+    t = (a->flags ^ b->flags) & condition & BN_FLG_CONSTTIME;
+    a->flags ^= t;
+    b->flags ^= t;
+
 #define BN_CONSTTIME_SWAP(ind) \
         do { \
                 t = (a->d[ind] ^ b->d[ind]) & condition; \

diff --git a/crypto/ec/ec_mult.c b/crypto/ec/ec_mult.c
index ed26b68c301ebcc1df1ebed3d2f7ef67f3d1ae44..1b9a4cf2ae179d249fed4ccb7801d4dc94cfb0b3 100644 (file)
--- a/crypto/ec/ec_mult.c
+++ b/crypto/ec/ec_mult.c
@@ -101,6 +101,166 @@ void EC_ec_pre_comp_free(EC_PRE_COMP *pre)
     OPENSSL_free(pre);
 }
 
+#define EC_POINT_set_flags(P, flags) do { \
+    BN_set_flags((P)->X, (flags)); \
+    BN_set_flags((P)->Y, (flags)); \
+    BN_set_flags((P)->Z, (flags)); \
+} while(0)
+
+/*
+ * This functions computes (in constant time) a point multiplication over the
+ * EC group.
+ *
+ * It performs either a fixed scalar point multiplication
+ *          (scalar * generator)
+ * when point is NULL, or a generic scalar point multiplication
+ *          (scalar * point)
+ * when point is not NULL.
+ *
+ * scalar should be in the range [0,n) otherwise all constant time bets are off.
+ *
+ * NB: This says nothing about EC_POINT_add and EC_POINT_dbl,
+ * which of course are not constant time themselves.
+ *
+ * The product is stored in r.
+ *
+ * Returns 1 on success, 0 otherwise.
+ */
+static int ec_mul_consttime(const EC_GROUP *group, EC_POINT *r, const BIGNUM *scalar,
+                            const EC_POINT *point, BN_CTX *ctx)
+{
+    int i, order_bits, group_top, kbit, pbit, Z_is_one, ret;
+    ret = 0;
+    EC_POINT *s = NULL;
+    BIGNUM *k = NULL;
+    BIGNUM *lambda = NULL;
+    BN_CTX *new_ctx = NULL;
+
+    if (ctx == NULL)
+        if ((ctx = new_ctx = BN_CTX_secure_new()) == NULL)
+            return 0;
+
+    if ((group->order == NULL) || (group->field == NULL))
+        goto err;
+
+    order_bits = BN_num_bits(group->order);
+
+    s = EC_POINT_new(group);
+    if (s == NULL)
+        goto err;
+
+    if (point == NULL) {
+        if (group->generator == NULL)
+            goto err;
+        if (!EC_POINT_copy(s, group->generator))
+            goto err;
+    } else {
+        if (!EC_POINT_copy(s, point))
+            goto err;
+    }
+
+    EC_POINT_set_flags(s, BN_FLG_CONSTTIME);
+
+    BN_CTX_start(ctx);
+    lambda = BN_CTX_get(ctx);
+    k = BN_CTX_get(ctx);
+    if (k == NULL)
+        goto err;
+
+    /*
+     * Group orders are often on a word boundary.
+     * So when we pad the scalar, some timing diff might
+     * pop if it needs to be expanded due to carries.
+     * So expand ahead of time.
+     */
+    group_top = bn_get_top(group->order);
+    if ((bn_wexpand(k, group_top + 1) == NULL)
+        || (bn_wexpand(lambda, group_top + 1) == NULL))
+        goto err;
+
+    if (!BN_copy(k, scalar))
+        goto err;
+
+    BN_set_flags(k, BN_FLG_CONSTTIME);
+
+    if ((BN_num_bits(k) > order_bits) || (BN_is_negative(k))) {
+        /*
+         * this is an unusual input, and we don't guarantee
+         * constant-timeness
+         */
+        if(!BN_nnmod(k, k, group->order, ctx))
+            goto err;
+    }
+
+    if (!BN_add(lambda, k, group->order))
+        goto err;
+    BN_set_flags(lambda, BN_FLG_CONSTTIME);
+    if (!BN_add(k, lambda, group->order))
+        goto err;
+    /*
+     * lambda := scalar + order
+     * k := scalar + 2*order
+     */
+    kbit = BN_is_bit_set(lambda, order_bits);
+    BN_consttime_swap(kbit, k, lambda, group_top + 1);
+
+    group_top = bn_get_top(group->field);
+    if ((bn_wexpand(s->X, group_top) == NULL)
+        || (bn_wexpand(s->Y, group_top) == NULL)
+        || (bn_wexpand(s->Z, group_top) == NULL)
+        || (bn_wexpand(r->X, group_top) == NULL)
+        || (bn_wexpand(r->Y, group_top) == NULL)
+        || (bn_wexpand(r->Z, group_top) == NULL))
+        goto err;
+
+    /* top bit is a 1, in a fixed pos */
+    if (!EC_POINT_copy(r, s))
+        goto err;
+
+    EC_POINT_set_flags(r, BN_FLG_CONSTTIME);
+
+    if (!EC_POINT_dbl(group, s, s, ctx))
+        goto err;
+
+    pbit = 0;
+
+#define EC_POINT_CSWAP(c, a, b, w, t) do {         \
+        BN_consttime_swap(c, (a)->X, (b)->X, w);   \
+        BN_consttime_swap(c, (a)->Y, (b)->Y, w);   \
+        BN_consttime_swap(c, (a)->Z, (b)->Z, w);   \
+        t = ((a)->Z_is_one ^ (b)->Z_is_one) & (c); \
+        (a)->Z_is_one ^= (t);                      \
+        (b)->Z_is_one ^= (t);                      \
+} while(0)
+
+    for (i = order_bits - 1; i >= 0; i--) {
+        kbit = BN_is_bit_set(k, i) ^ pbit;
+        EC_POINT_CSWAP(kbit, r, s, group_top, Z_is_one);
+        if (!EC_POINT_add(group, s, r, s, ctx))
+            goto err;
+        if (!EC_POINT_dbl(group, r, r, ctx))
+            goto err;
+        /*
+         * pbit logic merges this cswap with that of the
+         * next iteration
+         */
+        pbit ^= kbit;
+    }
+    /* one final cswap to move the right value into r */
+    EC_POINT_CSWAP(pbit, r, s, group_top, Z_is_one);
+#undef EC_POINT_CSWAP
+
+    ret = 1;
+
+err:
+    EC_POINT_free(s);
+    BN_CTX_end(ctx);
+    BN_CTX_free(new_ctx);
+
+    return ret;
+}
+#undef EC_POINT_set_flags
+
 /*
  * TODO: table should be optimised for the wNAF-based implementation,
  * sometimes smaller windows will give better performance (thus the
@@ -126,6 +286,28 @@ int ec_wNAF_mul(const EC_GROUP *group, EC_POINT *r, const BIGNUM *scalar,
                 size_t num, const EC_POINT *points[], const BIGNUM *scalars[],
                 BN_CTX *ctx)
 {
+    if ((scalar != NULL) && (num == 0)) {
+        /* In this case we want to compute scalar * GeneratorPoint:
+         * this codepath is reached most prominently by (ephemeral) key
+         * generation of EC cryptosystems (i.e. ECDSA keygen and sign setup,
+         * ECDH keygen/first half), where the scalar is always secret.
+         * This is why we ignore if BN_FLG_CONSTTIME is actually set and we
+         * always call the constant time version.
+         */
+        return ec_mul_consttime(group, r, scalar, NULL, ctx);
+    }
+
+    if ((scalar == NULL) && (num == 1)) {
+        /* In this case we want to compute scalar * GenericPoint:
+         * this codepath is reached most prominently by the second half of
+         * ECDH, where the secret scalar is multiplied by the peer's public
+         * point.
+         * To protect the secret scalar, we ignore if BN_FLG_CONSTTIME is
+         * actually set and we always call the constant time version.
+         */
+        return ec_mul_consttime(group, r, scalars[0], points[0], ctx);
+    }
+
     BN_CTX *new_ctx = NULL;
     const EC_POINT *generator = NULL;
     EC_POINT *tmp = NULL;