The clean way to use mbstowcs is to call it twice to find the length of the result:
const char * cs = <your input char*>
size_t wn = mbsrtowcs(NULL, &cs, 0, NULL);
// error if wn == size_t(-1)
wchar_t * buf = new wchar_t[wn + 1](); // value-initialize to 0 (see below)
wn = mbsrtowcs(buf, &cs, wn + 1, NULL);
// error if wn == size_t(-1)
assert(cs == NULL); // successful conversion
// result now in buf, return e.g. as std::wstring
delete[] buf;
Don't forget to call setlocale(LC_CTYPE, ""); at the beginning of your program!
The advantage over the Windows MultiByteToWideChar is that this is entirely standard C, although on Windows you might prefer the Windows API function anyway.
I usually wrap this method, along with the opposite one, in two conversion functions string->wstring and wstring->string. If you also add trivial overloads string->string and wstring->wstring, you can easily write code that compiles with the Winapi TCHAR typedef in any setting.
[Edit:] I added zero-initialization to buf, in case you plan to use the C array directly. I would usually return the result as std::wstring(buf, wn), though, but do beware if you plan on using C-style null-terminated arrays.[/]
In a multithreaded environment you should pass a thread-local conversion state to the function as its final (currently invisible) parameter.
Here is a small rant of mine on this topic.