On 03/20/14 20:21, Michael Roth wrote:
> Quoting Markus Armbruster (2014-03-18 04:32:08)
>> Peter Maydell <peter.mayd...@linaro.org> writes:
>>
>>> This is something clang's -fsanitize=undefined spotted. The
>>> code generated by qapi-commands.py in qmp-marshal.c for
>>> qmp_marshal_* functions where there are some optional
>>> arguments looks like this:
>>>
>>> bool has_force = false;
>>> bool force;
>>>
>>> mi = qmp_input_visitor_new_strict(QOBJECT(args));
>>> v = qmp_input_get_visitor(mi);
>>> visit_type_str(v, &device, "device", errp);
>>> visit_start_optional(v, &has_force, "force", errp);
>>> if (has_force) {
>>> visit_type_bool(v, &force, "force", errp);
>>> }
>>> visit_end_optional(v, errp);
>>> qmp_input_visitor_cleanup(mi);
>>>
>>> if (error_is_set(errp)) {
>>> goto out;
>>> }
>>> qmp_eject(device, has_force, force, errp);
>>>
>>> In the case where has_force is false, we never initialize
>>> force, but then we use it by passing it to qmp_eject.
>>> I imagine we don't then actually use the value, but clang
>>
>> Use of FOO when !has_FOO is a bug.
>>
>>> complains in particular for 'bool' variables because the value
>>> that ends up being loaded from memory for 'force' is not either
>>> 0 or 1 (being uninitialized stack contents).
>>>
>>> Anybody understand what the codegenerator is doing well enough
>>> to suggest a fix? I'd guess that just initializing the variable either
>>> at point of declaration or in an else {) clause of the 'if (has_force)'
>>> conditional would suffice, but presumably you need to handle
>>> all the possible data types...
>>
>> I can give it a try. Will probably take a while, though.
>
> Could it be as simple as this?:
>
> diff --git a/scripts/qapi-commands.py b/scripts/qapi-commands.py
> index 9734ab0..a70482e 100644
> --- a/scripts/qapi-commands.py
> +++ b/scripts/qapi-commands.py
> @@ -99,7 +99,7 @@ bool has_%(argname)s = false;
> argname=c_var(argname), argtype=c_type(argtype))
> else:
> ret += mcgen('''
> -%(argtype)s %(argname)s;
> +%(argtype)s %(argname)s = {0};
> ''',
> argname=c_var(argname), argtype=c_type(argtype))
>
> Pointer-type are special-cased initialized to NULL, so that leaves these guys
> in the current set of qapi-defined types that we use as direct arguments for
> qmp commands:
>
> NON-POINTER TYPE: BlockdevOnError
> NON-POINTER TYPE: bool
> NON-POINTER TYPE: DataFormat
> NON-POINTER TYPE: double
> NON-POINTER TYPE: DumpGuestMemoryFormat
> NON-POINTER TYPE: int64_t
> NON-POINTER TYPE: MirrorSyncMode
> NON-POINTER TYPE: NewImageMode
> NON-POINTER TYPE: uint32_t
>
> I'm trying to make sense of whether {0} is a valid initializer in all these
> cases, as I saw some references to GCC complaining about cases where you don't
> use an initializer for each nested subtype (back in 2002 at least:
> http://www.ex-parrot.com/~chris/random/initialise.html), but that doesn't seem
> to be the case now.
>
> If that's not safe, we can memset based on sizeof() in the else clause, but
> obviously that's sub-optimal.
{ 0 } is safe. { 0 } is a "universal initializer". If you tell me which
C version we care about this week, I can look up and cite the language
for you. The gist, as far as I remember, is that
- 0 is a good initializer for any scalar type,
- the outermost braces are ignored when initializing a scalar,
- the outermost braces allow initialization of an aggregate (struct or
array) or a union,
- sub-aggregates don't require further braces.
Thanks,
Laszlo
