Re: [Qemu-devel] Error propagation in generated visitors and command marshallers

[Peter Crosthwaite]

On Thu, Apr 10, 2014 at 1:48 AM, Markus Armbruster <address@hidden> wrote:
> I stumbled over this while trying to purge error_is_set() from the code.
>
>
> Here's how we commonly use the Error API:
>
>     Error *err = NULL;
>
>     foo(arg, &err)
>     if (err) {
>         goto out;
>     }
>     bar(arg, &err)
>     if (err) {
>         goto out;
>     }
>
> This ensures that err is null on entry, both for foo() and for bar().
> Many functions rely on that, like this:
>
>     void foo(ArgType arg, Error **errp)
>     {
>         if (frobnicate(arg) < 0) {
>             error_setg(errp, "Can't frobnicate");
>                                 // This asserts errp != NULL
>         }
>     }
>
>
> Here's how some of our visitor code uses the Error API (for real code,
> check out generated qmp-marshal.c):
>
>     Error *err = NULL;
>     QmpInputVisitor *mi = qmp_input_visitor_new_strict(QOBJECT(args));
>     Visitor *v = qmp_input_get_visitor(mi);
>     char *foo = NULL;
>     char *bar = NULL;
>
>     visit_type_str(v, &foo, "foo", &err);
>     visit_type_str(v, &bar, "bar", &err);
>     if (err) {
>         goto out;
>     }
>
> Unlike above, this may pass a non-null errp to the second
> visit_type_str(), namely when the first one fails.
>
> The visitor functions guard against that, like this:
>
>     void visit_type_str(Visitor *v, char **obj, const char *name, Error 
> **errp)
>     {
>         if (!error_is_set(errp)) {
>             v->type_str(v, obj, name, errp);
>         }
>     }
>
> As discussed before, error_is_set() is almost almost wrong, fragile or
> unclean.  What if errp is null?  Then we fail to stop visiting after an
> error.

That should be the callers problem. If you pass a NULL errp then the
intended semantic is to ignore errors.

>
> The function could be improved like this:
>
>     void visit_type_str(Visitor *v, char **obj, const char *name, Error 
> **errp)
>     {
>         assert(errp);

And this is irregular in that you are now mandating the Error ** and
thus removing the capability to ignore errors.

>         if (!*errp) {
>             v->type_str(v, obj, name, errp);
>         }
>     }
>
>
> But: is it a good idea to have both patterns in the code?  Should we
> perhaps use the common pattern for visiting, too?  Like this:
>
>     visit_type_str(v, &foo, "foo", &err);
>     if (err) {
>         goto out;
>     }
>     visit_type_str(v, &bar, "bar", &err);
>     if (err) {
>         goto out;
>     }
>
> Then we can assume *errp is clear on function entry, like this:
>
>     void visit_type_str(Visitor *v, char **obj, const char *name, Error 
> **errp)
>     {
>         v->type_str(v, obj, name, errp);
>     }
>
> Should execute roughly the same number of conditional branches.
>
> Tedious repetition of "if (err) goto out" in the caller, but that's what
> we do elsewhere, and unlike elsewhere, these one's are generated.
>
> Opinions?

I think this code as-is is a good example of what we should do
elsewhere. The code base has bloated with the if (error) { bail; } on
every Error ** accepting API call. I proposed a while back a semantic
that Error ** Accepting APIs perform no action when the error is
already set to allow for long sequences of calls to run without the
constant checks. You then report the first error in a catchall at the
end of the run.

I think this particular code is probably good, provided your case of
NULL errp is enforced against by the caller.

Regards,
Peter