Blame SOURCES/openssl-1.0.2k-cve-2018-5407.patch

a07c00
diff -up openssl-1.0.2k/crypto/bn/bn_lib.c.ecc-ladder openssl-1.0.2k/crypto/bn/bn_lib.c
a07c00
--- openssl-1.0.2k/crypto/bn/bn_lib.c.ecc-ladder	2019-02-06 12:58:50.575844123 +0100
a07c00
+++ openssl-1.0.2k/crypto/bn/bn_lib.c	2019-02-08 10:48:53.529291777 +0100
a07c00
@@ -877,6 +877,38 @@ void BN_consttime_swap(BN_ULONG conditio
a07c00
     a->top ^= t;
a07c00
     b->top ^= t;
a07c00
 
a07c00
+    t = (a->neg ^ b->neg) & condition;
a07c00
+    a->neg ^= t;
a07c00
+    b->neg ^= t;
a07c00
+
a07c00
+    /*-
a07c00
+     * BN_FLG_STATIC_DATA: indicates that data may not be written to. Intention
a07c00
+     * is actually to treat it as it's read-only data, and some (if not most)
a07c00
+     * of it does reside in read-only segment. In other words observation of
a07c00
+     * BN_FLG_STATIC_DATA in BN_consttime_swap should be treated as fatal
a07c00
+     * condition. It would either cause SEGV or effectively cause data
a07c00
+     * corruption.
a07c00
+     *
a07c00
+     * BN_FLG_MALLOCED: refers to BN structure itself, and hence must be
a07c00
+     * preserved.
a07c00
+     *
a07c00
+     * BN_FLG_SECURE: must be preserved, because it determines how x->d was
a07c00
+     * allocated and hence how to free it.
a07c00
+     *
a07c00
+     * BN_FLG_CONSTTIME: sufficient to mask and swap
a07c00
+     *
a07c00
+     * BN_FLG_FIXED_TOP: indicates that we haven't called bn_correct_top() on
a07c00
+     * the data, so the d array may be padded with additional 0 values (i.e.
a07c00
+     * top could be greater than the minimal value that it could be). We should
a07c00
+     * be swapping it
a07c00
+     */
a07c00
+
a07c00
+#define BN_CONSTTIME_SWAP_FLAGS (BN_FLG_CONSTTIME | BN_FLG_FIXED_TOP)
a07c00
+
a07c00
+    t = ((a->flags ^ b->flags) & BN_CONSTTIME_SWAP_FLAGS) & condition;
a07c00
+    a->flags ^= t;
a07c00
+    b->flags ^= t;
a07c00
+
a07c00
 #define BN_CONSTTIME_SWAP(ind) \
a07c00
         do { \
a07c00
                 t = (a->d[ind] ^ b->d[ind]) & condition; \
a07c00
diff -up openssl-1.0.2k/crypto/ec/ec_mult.c.ecc-ladder openssl-1.0.2k/crypto/ec/ec_mult.c
a07c00
--- openssl-1.0.2k/crypto/ec/ec_mult.c.ecc-ladder	2017-01-26 14:22:03.000000000 +0100
a07c00
+++ openssl-1.0.2k/crypto/ec/ec_mult.c	2019-02-08 10:48:53.531291744 +0100
a07c00
@@ -306,6 +306,224 @@ static signed char *compute_wNAF(const B
a07c00
     return r;
a07c00
 }
a07c00
 
a07c00
+#define EC_POINT_BN_set_flags(P, flags) do { \
a07c00
+    BN_set_flags(&(P)->X, (flags)); \
a07c00
+    BN_set_flags(&(P)->Y, (flags)); \
a07c00
+    BN_set_flags(&(P)->Z, (flags)); \
a07c00
+} while(0)
a07c00
+
a07c00
+/*-
a07c00
+ * This functions computes (in constant time) a point multiplication over the
a07c00
+ * EC group.
a07c00
+ *
a07c00
+ * At a high level, it is Montgomery ladder with conditional swaps.
a07c00
+ *
a07c00
+ * It performs either a fixed scalar point multiplication
a07c00
+ *          (scalar * generator)
a07c00
+ * when point is NULL, or a generic scalar point multiplication
a07c00
+ *          (scalar * point)
a07c00
+ * when point is not NULL.
a07c00
+ *
a07c00
+ * scalar should be in the range [0,n) otherwise all constant time bets are off.
a07c00
+ *
a07c00
+ * NB: This says nothing about EC_POINT_add and EC_POINT_dbl,
a07c00
+ * which of course are not constant time themselves.
a07c00
+ *
a07c00
+ * The product is stored in r.
a07c00
+ *
a07c00
+ * Returns 1 on success, 0 otherwise.
a07c00
+ */
a07c00
+static int ec_mul_consttime(const EC_GROUP *group, EC_POINT *r,
a07c00
+                            const BIGNUM *scalar, const EC_POINT *point,
a07c00
+                            BN_CTX *ctx)
a07c00
+{
a07c00
+    int i, cardinality_bits, group_top, kbit, pbit, Z_is_one;
a07c00
+    EC_POINT *s = NULL;
a07c00
+    BIGNUM *k = NULL;
a07c00
+    BIGNUM *lambda = NULL;
a07c00
+    BIGNUM *cardinality = NULL;
a07c00
+    BN_CTX *new_ctx = NULL;
a07c00
+    int ret = 0;
a07c00
+
a07c00
+    if (ctx == NULL && (ctx = new_ctx = BN_CTX_new()) == NULL)
a07c00
+        return 0;
a07c00
+
a07c00
+    BN_CTX_start(ctx);
a07c00
+
a07c00
+    s = EC_POINT_new(group);
a07c00
+    if (s == NULL)
a07c00
+        goto err;
a07c00
+
a07c00
+    if (point == NULL) {
a07c00
+        if (!EC_POINT_copy(s, group->generator))
a07c00
+            goto err;
a07c00
+    } else {
a07c00
+        if (!EC_POINT_copy(s, point))
a07c00
+            goto err;
a07c00
+    }
a07c00
+
a07c00
+    EC_POINT_BN_set_flags(s, BN_FLG_CONSTTIME);
a07c00
+
a07c00
+    cardinality = BN_CTX_get(ctx);
a07c00
+    lambda = BN_CTX_get(ctx);
a07c00
+    k = BN_CTX_get(ctx);
a07c00
+    if (k == NULL || !BN_mul(cardinality, &group->order, &group->cofactor, ctx))
a07c00
+        goto err;
a07c00
+
a07c00
+    /*
a07c00
+     * Group cardinalities are often on a word boundary.
a07c00
+     * So when we pad the scalar, some timing diff might
a07c00
+     * pop if it needs to be expanded due to carries.
a07c00
+     * So expand ahead of time.
a07c00
+     */
a07c00
+    cardinality_bits = BN_num_bits(cardinality);
a07c00
+    group_top = cardinality->top;
a07c00
+    if ((bn_wexpand(k, group_top + 2) == NULL)
a07c00
+        || (bn_wexpand(lambda, group_top + 2) == NULL))
a07c00
+        goto err;
a07c00
+
a07c00
+    if (!BN_copy(k, scalar))
a07c00
+        goto err;
a07c00
+
a07c00
+    BN_set_flags(k, BN_FLG_CONSTTIME);
a07c00
+
a07c00
+    if ((BN_num_bits(k) > cardinality_bits) || (BN_is_negative(k))) {
a07c00
+        /*-
a07c00
+         * this is an unusual input, and we don't guarantee
a07c00
+         * constant-timeness
a07c00
+         */
a07c00
+        if (!BN_nnmod(k, k, cardinality, ctx))
a07c00
+            goto err;
a07c00
+    }
a07c00
+
a07c00
+    if (!BN_add(lambda, k, cardinality))
a07c00
+        goto err;
a07c00
+    BN_set_flags(lambda, BN_FLG_CONSTTIME);
a07c00
+    if (!BN_add(k, lambda, cardinality))
a07c00
+        goto err;
a07c00
+    /*
a07c00
+     * lambda := scalar + cardinality
a07c00
+     * k := scalar + 2*cardinality
a07c00
+     */
a07c00
+    kbit = BN_is_bit_set(lambda, cardinality_bits);
a07c00
+    BN_consttime_swap(kbit, k, lambda, group_top + 2);
a07c00
+
a07c00
+    group_top = group->field.top;
a07c00
+    if ((bn_wexpand(&s->X, group_top) == NULL)
a07c00
+        || (bn_wexpand(&s->Y, group_top) == NULL)
a07c00
+        || (bn_wexpand(&s->Z, group_top) == NULL)
a07c00
+        || (bn_wexpand(&r->X, group_top) == NULL)
a07c00
+        || (bn_wexpand(&r->Y, group_top) == NULL)
a07c00
+        || (bn_wexpand(&r->Z, group_top) == NULL))
a07c00
+        goto err;
a07c00
+
a07c00
+    /* top bit is a 1, in a fixed pos */
a07c00
+    if (!EC_POINT_copy(r, s))
a07c00
+        goto err;
a07c00
+
a07c00
+    EC_POINT_BN_set_flags(r, BN_FLG_CONSTTIME);
a07c00
+
a07c00
+    if (!EC_POINT_dbl(group, s, s, ctx))
a07c00
+        goto err;
a07c00
+
a07c00
+    pbit = 0;
a07c00
+
a07c00
+#define EC_POINT_CSWAP(c, a, b, w, t) do {         \
a07c00
+        BN_consttime_swap(c, &(a)->X, &(b)->X, w); \
a07c00
+        BN_consttime_swap(c, &(a)->Y, &(b)->Y, w); \
a07c00
+        BN_consttime_swap(c, &(a)->Z, &(b)->Z, w); \
a07c00
+        t = ((a)->Z_is_one ^ (b)->Z_is_one) & (c); \
a07c00
+        (a)->Z_is_one ^= (t);                      \
a07c00
+        (b)->Z_is_one ^= (t);                      \
a07c00
+} while(0)
a07c00
+
a07c00
+    /*-
a07c00
+     * The ladder step, with branches, is
a07c00
+     *
a07c00
+     * k[i] == 0: S = add(R, S), R = dbl(R)
a07c00
+     * k[i] == 1: R = add(S, R), S = dbl(S)
a07c00
+     *
a07c00
+     * Swapping R, S conditionally on k[i] leaves you with state
a07c00
+     *
a07c00
+     * k[i] == 0: T, U = R, S
a07c00
+     * k[i] == 1: T, U = S, R
a07c00
+     *
a07c00
+     * Then perform the ECC ops.
a07c00
+     *
a07c00
+     * U = add(T, U)
a07c00
+     * T = dbl(T)
a07c00
+     *
a07c00
+     * Which leaves you with state
a07c00
+     *
a07c00
+     * k[i] == 0: U = add(R, S), T = dbl(R)
a07c00
+     * k[i] == 1: U = add(S, R), T = dbl(S)
a07c00
+     *
a07c00
+     * Swapping T, U conditionally on k[i] leaves you with state
a07c00
+     *
a07c00
+     * k[i] == 0: R, S = T, U
a07c00
+     * k[i] == 1: R, S = U, T
a07c00
+     *
a07c00
+     * Which leaves you with state
a07c00
+     *
a07c00
+     * k[i] == 0: S = add(R, S), R = dbl(R)
a07c00
+     * k[i] == 1: R = add(S, R), S = dbl(S)
a07c00
+     *
a07c00
+     * So we get the same logic, but instead of a branch it's a
a07c00
+     * conditional swap, followed by ECC ops, then another conditional swap.
a07c00
+     *
a07c00
+     * Optimization: The end of iteration i and start of i-1 looks like
a07c00
+     *
a07c00
+     * ...
a07c00
+     * CSWAP(k[i], R, S)
a07c00
+     * ECC
a07c00
+     * CSWAP(k[i], R, S)
a07c00
+     * (next iteration)
a07c00
+     * CSWAP(k[i-1], R, S)
a07c00
+     * ECC
a07c00
+     * CSWAP(k[i-1], R, S)
a07c00
+     * ...
a07c00
+     *
a07c00
+     * So instead of two contiguous swaps, you can merge the condition
a07c00
+     * bits and do a single swap.
a07c00
+     *
a07c00
+     * k[i]   k[i-1]    Outcome
a07c00
+     * 0      0         No Swap
a07c00
+     * 0      1         Swap
a07c00
+     * 1      0         Swap
a07c00
+     * 1      1         No Swap
a07c00
+     *
a07c00
+     * This is XOR. pbit tracks the previous bit of k.
a07c00
+     */
a07c00
+
a07c00
+    for (i = cardinality_bits - 1; i >= 0; i--) {
a07c00
+        kbit = BN_is_bit_set(k, i) ^ pbit;
a07c00
+        EC_POINT_CSWAP(kbit, r, s, group_top, Z_is_one);
a07c00
+        if (!EC_POINT_add(group, s, r, s, ctx))
a07c00
+            goto err;
a07c00
+        if (!EC_POINT_dbl(group, r, r, ctx))
a07c00
+            goto err;
a07c00
+        /*
a07c00
+         * pbit logic merges this cswap with that of the
a07c00
+         * next iteration
a07c00
+         */
a07c00
+        pbit ^= kbit;
a07c00
+    }
a07c00
+    /* one final cswap to move the right value into r */
a07c00
+    EC_POINT_CSWAP(pbit, r, s, group_top, Z_is_one);
a07c00
+#undef EC_POINT_CSWAP
a07c00
+
a07c00
+    ret = 1;
a07c00
+
a07c00
+ err:
a07c00
+    EC_POINT_free(s);
a07c00
+    BN_CTX_end(ctx);
a07c00
+    BN_CTX_free(new_ctx);
a07c00
+
a07c00
+    return ret;
a07c00
+}
a07c00
+
a07c00
+#undef EC_POINT_BN_set_flags
a07c00
+
a07c00
 /*
a07c00
  * TODO: table should be optimised for the wNAF-based implementation,
a07c00
  * sometimes smaller windows will give better performance (thus the
a07c00
@@ -365,6 +583,34 @@ int ec_wNAF_mul(const EC_GROUP *group, E
a07c00
         return EC_POINT_set_to_infinity(group, r);
a07c00
     }
a07c00
 
a07c00
+    if (!BN_is_zero(&group->order) && !BN_is_zero(&group->cofactor)) {
a07c00
+        /*-
a07c00
+         * Handle the common cases where the scalar is secret, enforcing a constant
a07c00
+         * time scalar multiplication algorithm.
a07c00
+         */
a07c00
+        if ((scalar != NULL) && (num == 0)) {
a07c00
+            /*-
a07c00
+             * In this case we want to compute scalar * GeneratorPoint: this
a07c00
+             * codepath is reached most prominently by (ephemeral) key generation
a07c00
+             * of EC cryptosystems (i.e. ECDSA keygen and sign setup, ECDH
a07c00
+             * keygen/first half), where the scalar is always secret. This is why
a07c00
+             * we ignore if BN_FLG_CONSTTIME is actually set and we always call the
a07c00
+             * constant time version.
a07c00
+             */
a07c00
+            return ec_mul_consttime(group, r, scalar, NULL, ctx);
a07c00
+        }
a07c00
+        if ((scalar == NULL) && (num == 1)) {
a07c00
+            /*-
a07c00
+             * In this case we want to compute scalar * GenericPoint: this codepath
a07c00
+             * is reached most prominently by the second half of ECDH, where the
a07c00
+             * secret scalar is multiplied by the peer's public point. To protect
a07c00
+             * the secret scalar, we ignore if BN_FLG_CONSTTIME is actually set and
a07c00
+             * we always call the constant time version.
a07c00
+             */
a07c00
+            return ec_mul_consttime(group, r, scalars[0], points[0], ctx);
a07c00
+        }
a07c00
+    }
a07c00
+
a07c00
     for (i = 0; i < num; i++) {
a07c00
         if (group->meth != points[i]->meth) {
a07c00
             ECerr(EC_F_EC_WNAF_MUL, EC_R_INCOMPATIBLE_OBJECTS);