Re: [Qemu-devel] [PATCH v2] make check-unit: use after free in test-opts-visitor

[Markus Armbruster]

Andrey Shinkevich <address@hidden> writes:

> On 02/08/2019 14:34, Markus Armbruster wrote:
>> Andrey Shinkevich <address@hidden> writes:
>> 
>>> In struct OptsVisitor, repeated_opts member points to a list in the
>>> unprocessed_opts hash table after the list has been destroyed. A
>>> subsequent call to visit_type_int() references the deleted list. It
>>> results in use-after-free issue.
>> 
>> Let's mention the reproducer: valgrind tests/test/opts-visitor.
>> 
>>>                                   Also, the Visitor object call back
>>> functions are supposed to set the Error parameter in case of failure.
>> 
>> As far as I can tell, they all do.  The only place where you set an
>> error is the new failure you add to lookup_scalar().
>> 
>
> The story behind the comment is that the original 
> tests/test-opts-visitor fails being run under the Valgrind with the 
> error message:
>
> test-opts-visitor: util/error.c:276: error_free_or_abort: Assertion 
> `errp && *errp' failed.
>
> coming from
>
> assert(errp && *errp);
> error_free_or_abort (util/error.c:276)
> test_opts_range_beyond (tests/test-opts-visitor.c:241)
>
> because g_queue_peek_head() returns NULL under the Valgrind and errp 
> stays unset.
>
> Without the Valgrind, the g_queue_peek_head() returns a non-zero pointer 
> and the opts_type_int64() sets the following error:
>
> "Parameter '\340F\212\274\267U' expects an int64 value or range", 
> err_class = ERROR_CLASS_GENERIC_ERROR, src = 0x55b7bbc02163 
> "qapi/opts-visitor.c",
>    func = 0x55b7bbc02410 <__func__.14916> "opts_type_int64", line = 433, 
> hint = 0x0}
>
> so, error_free_or_abort() doesn't abort and the test case passes.
>
> I will remove the comment in v3.

Thanks.

[...]
>>> diff --git a/qapi/opts-visitor.c b/qapi/opts-visitor.c
>>> index 324b197..23ac383 100644
>>> --- a/qapi/opts-visitor.c
>>> +++ b/qapi/opts-visitor.c
[...]
>>> @@ -289,8 +302,11 @@ opts_end_list(Visitor *v, void **obj)
>>>   
>>>       assert(ov->list_mode == LM_IN_PROGRESS ||
>>>              ov->list_mode == LM_SIGNED_INTERVAL ||
>>> -           ov->list_mode == LM_UNSIGNED_INTERVAL);
>>> -    ov->repeated_opts = NULL;
>>> +           ov->list_mode == LM_UNSIGNED_INTERVAL ||
>>> +           ov->list_mode == LM_TRAVERSED);
>>> +    if (ov->list_mode != LM_TRAVERSED) {
>>> +        ov->repeated_opts = NULL;
>>> +    }
>> 
>> What's wrong with zapping ov->repeated_opts unconditionally?
>> 
>>>       ov->list_mode = LM_NONE;
>>>   }
>>>   
>>> @@ -306,6 +322,10 @@ lookup_scalar(const OptsVisitor *ov, const char *name, 
>>> Error **errp)
>>>           list = lookup_distinct(ov, name, errp);
>>>           return list ? g_queue_peek_tail(list) : NULL;
>>>       }
>>> +    if (ov->list_mode == LM_TRAVERSED) {
>>> +        error_setg(errp, QERR_INVALID_PARAMETER, name);
>> 
>> Beware, @name is null when visiting list members.  The test still passes
>> for me, since g_strdup_vprintf() formats a null argument to %s as
>> "(null)".
>> 
>> For what it's worth, the qobject input visitor uses
>> QERR_MISSING_PARAMETER with a made-up name.  Computing the name is
>> pretty elaborate, see full_name_nth().  I'd rather not duplicate that
>> here.
>> 
>> Suggest something like
>> 
>>             error_setg(errp, "Fewer list elements than expected");
>> 
>> The error message fails to mention the name of the list.  Bad, but the
>> error is a corner case; we don't normally visit beyond the end of the
>> list.  For a better message, we'd have to have start_list() store its
>> @name in struct OptsVisitor.  I'm not asking you to do that now.
>> 
>
> Will I do that work to add the @name of the list to the struct 
> OptsVisitor in a following series?

Entirely up to you.  To be honest, I'm not sure it's worth the trouble.

>>> +        return NULL;
>>> +    }
>>>       assert(ov->list_mode == LM_IN_PROGRESS);
>>>       return g_queue_peek_head(ov->repeated_opts);
>>>   }
>> 
>> I checked the remaining uses of ->list_mode, and I think they are okay.
>> 
> Thank you. I also had not noticed any potential issue with the list_mode.

Good :)