Implement cx_bn_gf2_n_mul

Workaround that implements cx_bn_gf2_n_mul() in the SDK until it is hopefully added as a syscall like all other Ledger devices
Fixes LedgerHQ#64

include/ox_bn.h

@@ -1075,4 +1075,30 @@ SYSCALL cx_err_t cx_bn_next_prime(cx_bn_t n);
  */
 SYSCALL cx_err_t cx_bn_rng(cx_bn_t r, const cx_bn_t n);
 
+/**
+ * @brief Performs a multiplication over GF(2^n).
+ *
+ * @param[out] bn_r BN index for the result.
+ *
+ * @param[in]  bn_a BN index of the first operand.
+ *
+ * @param[in]  bn_b BN index of the second operand.
+ *
+ * @param[in]  bn_n BN index of the modulus.
+ *                  The modulus must be an irreducible polynomial over GF(2)
+ *                  of degree n.
+ *
+ * @param[in]  bn_h BN index of the second montgomery constant.
+ *
+ * @return          Error code:
+ *                  - CX_OK on success
+ *                  - CX_NOT_LOCKED
+ *                  - CX_INVALID_PARAMETER
+ *                  - CX_MEMORY_FULL
+ */
+WARN_UNUSED_RESULT cx_err_t cx_bn_gf2_n_mul(cx_bn_t       bn_r,
+                                            const cx_bn_t bn_a,
+                                            const cx_bn_t bn_b,
+                                            const cx_bn_t bn_n,
+                                            const cx_bn_t bn_h);
 #endif /* CX_BN_H */

src/cx_wrappers.c

@@ -113,3 +113,72 @@ uint32_t cx_crc32(const void *buf, size_t len)
 {
     return cx_crc32_update(CX_CRC32_INIT, buf, len);
 }
+
+// At some point this could/should be migrated to the OS
+// Meanwhile, it's implemented here
+cx_err_t cx_bn_gf2_n_mul(cx_bn_t bn_r,
+                         const cx_bn_t bn_a,
+                         const cx_bn_t bn_b,
+                         const cx_bn_t bn_n,
+                         const cx_bn_t bn_h __attribute__((unused))) {
+
+    cx_err_t error = CX_OK;
+    uint32_t degree, nbits_a, nbits_b;
+
+    // Calculate the degree of the modulus polynomial
+    CX_CHECK(cx_bn_cnt_bits(bn_n, &degree));
+    degree--;
+
+    CX_CHECK(cx_bn_cnt_bits(bn_a, &nbits_a));
+    CX_CHECK(cx_bn_cnt_bits(bn_b, &nbits_b));
+
+    // Ensure both operands are in field
+    if (degree < 1 || nbits_a > degree || nbits_b > degree) {
+        error = CX_INVALID_PARAMETER;
+        goto end;
+    }
+
+    // Preliminaries
+    cx_bn_t bn_tempa, bn_tempb, bn_tempx;
+    uint32_t bit_indexb = 0;
+    size_t nbytes;
+    bool bit_set = 0;
+
+    CX_CHECK(cx_bn_nbytes(bn_n, &nbytes));
+    CX_CHECK(cx_bn_alloc(&bn_tempa, nbytes));
+    CX_CHECK(cx_bn_alloc(&bn_tempb, nbytes));
+    CX_CHECK(cx_bn_alloc(&bn_tempx, nbytes));
+
+    CX_CHECK(cx_bn_copy(bn_tempa, bn_a));
+    CX_CHECK(cx_bn_copy(bn_tempb, bn_b));
+    CX_CHECK(cx_bn_set_u32(bn_r, (uint32_t) 0));
+
+    // Main loop for multiplication
+    if (nbits_a) {
+        while (nbits_b > bit_indexb) {
+            CX_CHECK(cx_bn_tst_bit(bn_tempb, bit_indexb, &bit_set));
+            if (bit_set) {
+                CX_CHECK(cx_bn_copy(bn_tempx, bn_r));
+                CX_CHECK(cx_bn_xor(bn_r, bn_tempa, bn_tempx));
+            }
+
+            CX_CHECK(cx_bn_shl(bn_tempa, 1));
+            CX_CHECK(cx_bn_tst_bit(bn_tempa, degree, &bit_set));
+
+            if (bit_set) {
+                CX_CHECK(cx_bn_copy(bn_tempx, bn_tempa));
+                CX_CHECK(cx_bn_xor(bn_tempa, bn_n, bn_tempx));
+            }
+
+            bit_indexb++;
+        }
+    }
+
+    // Clean up
+    CX_CHECK(cx_bn_destroy(&bn_tempa));
+    CX_CHECK(cx_bn_destroy(&bn_tempb));
+    CX_CHECK(cx_bn_destroy(&bn_tempx));
+
+end:
+    return error;
+}