Re: [Qemu-devel] [PATCH v10 12/13] qapi: Delete unused visit_start_union()

[Markus Armbruster]

Eric Blake <address@hidden> writes:

> On 02/17/2016 11:08 AM, Markus Armbruster wrote:
>> Eric Blake <address@hidden> writes:
>> 
>>> Commit cee2dedb noticed that if you have a partial flat union
>>> (such as if an input parse failed due to a missing
>>> discriminator), calling the dealloc visitor could result in
>>> trying to dereference a NULL pointer if we attempted to visit
>>> an object branch without an earlier successful call to
>>> visit_start_implicit_struct() allocating the pointer for that
>>> branch. But the "fix" it implemented requires the use of a
>>> '.data' member in the union, which may or may not be the same
>>> size as other branches of the union (consider a 32-bit platform
>>> where one of the branches is an int64), which feels fairly dirty.
>> 
>> Well, until the previous commit, it was the same, wasn't it?  All
>> pointers.
>
> For simple unions, you could have (well, still can have, until my later
> patch gets rid of the simple_union_type() magic):
>
> struct SU {
>     SUKind type;
>     union {
>         void *data;
>         int8_t byte;
>     } u;
> };

Begs the question why that works :)

> But you're right - for flat unions, ALL branches were represented as
> pointers (until this series unboxed them).
>
>> 
>>> Plus, as mentioned in that commit, it only works if you can
>>> assume that '.data' would be zero-initialized even if '.kind' was
>>> uninitialized, which is rather poor logic: our usage of
>>> visit_start_struct() happens to zero-initialize both fields,
>>> which means '.kind' is never truly uninitialized - but if we
>>> changed visit_start_struct() to use g_new() instead of g_new0(),
>>> then '.data' would not be any more reliable as a condition on
>>> whether to visit the branch matching '.kind', regardless of
>>> whether '.kind' was 0).
>>>
>>> Menawhile, now that we have just inlined the fields of all flat
>
> Meanwhile,
>
>>> unions, there is no longer the possibility of a null pointer to
>>> dereference in the first place.  Where the branch structure used
>>> to be separately allocated by visit_start_implicit_struct(), it
>>> is now just pointing to a subset of the memory already
>>> zero-allocated by visit_start_struct().
>
> I guess I may try and reword this slightly, and point to the fact that
> the NULL dereference was due to calling visit_start_implicit_FOO() (only
> done for flat unions; for simple unions the branches call
> visit_type_FOO(), and that call safely handled NULL);

That's why it works?

>                                                       because we were
> using visit_start/end_implicit_struct() for its allocation effects.  But
> the net result is the same - now that we no longer call
> visit_start_implicit_struct() for a union visit, the dealloc visitor no
> longer has to worry about a NULL dereference on a partially constructed
> object, so we no longer need to probe if the union contains any data.
>
>>> +++ b/scripts/qapi-visit.py
>>> @@ -246,9 +246,6 @@ void visit_type_%(c_name)s(Visitor *v, const char 
>>> *name, %(c_name)s **obj, Error
>>>      if variants:
>>>          ret += gen_err_check(label='out_obj')
>>>          ret += mcgen('''
>>> -    if (!visit_start_union(v, !!(*obj)->u.data, &err) || err) {
>>> -        goto out_obj;
>>> -    }
>> 
>> I'm afraid the previous commit broke this for flat unions.
>> 
>> Before the previous commit, all members of (*obj)->u were pointers to
>> the struct holding the variant members both for flat and simple unions.
>> !!(*obj)->u.data tests whether the struct holding the variant members
>> has been allocated.  This relies on uniform pointer format.
>> 
>> The dealloc visitor uses the "has been allocated" bit to suppress
>> visiting the struct when it hasn't been allocated.
>> 
>> The previous commit unboxes the struct for flat unions.  Now ->u.data
>> reinterprets the first few bytes of that struct as pointer.  If you're
>> "lucky", they're not all zero, and the struct gets visited.
>
> You're right - and I bet I could come up with a case where valgrind
> could call me on it.
>
>> 
>> Obvious fix: squash this hunk into the previous commit, then let this
>> commit drop the code that's no longer used.
>
> Yep, for bisectability, I think that's what I'll end up doing.
>
>> 
>> However, simple unions are still boxed.  Why can't their pointer be null
>> in the dealloc visitor?
>
> Simple unions still go through visit_type_FOO(), and _that_ function
> properly checks for NULL.  It was only visit_type_implicit_FOO() that
> blindly dereferenced things.  In fact, in the earlier incantation of
> this patch, my fix was to teach visit_type_implicit_FOO() how to check
> for NULL:
> https://lists.gnu.org/archive/html/qemu-devel/2015-09/msg05442.html
>
> But now that visit_type_implicit_FOO() is gone, my earlier incantation
> got reduced in size.  I guess it's all in how I document the commit message.

Give it a try :)