Re: [PATCH v8 09/20] kasan: preassign tags to objects with ctors or SLAB_TYPESAFE_BY_RCU

[Dmitry Vyukov <dvyukov@google.com>]

On Fri, Sep 21, 2018 at 2:24 PM, Andrey Konovalov <andreyknvl@google.com> wrote:
> On Fri, Sep 21, 2018 at 1:25 PM, Dmitry Vyukov <dvyukov@google.com> wrote:
>> On Wed, Sep 19, 2018 at 8:54 PM, Andrey Konovalov <andreyknvl@google.com> wrote:
>
>>>         if (!shuffle) {
>>>                 for_each_object_idx(p, idx, s, start, page->objects) {
>>> -                       setup_object(s, page, p);
>>> -                       if (likely(idx < page->objects))
>>> -                               set_freepointer(s, p, p + s->size);
>>> -                       else
>>> +                       if (likely(idx < page->objects)) {
>>> +                               next = p + s->size;
>>> +                               next = setup_object(s, page, next);
>>> +                               set_freepointer(s, p, next);
>>> +                       } else
>>>                                 set_freepointer(s, p, NULL);
>>>                 }
>>> -               page->freelist = fixup_red_left(s, start);
>>> +               start = fixup_red_left(s, start);
>>> +               start = setup_object(s, page, start);
>>> +               page->freelist = start;
>>>         }
>>
>> Just want to double-check that this is correct.
>> We now do an additional setup_object call after the loop, but we do 1
>> less in the loop. So total number of calls should be the same, right?
>> However, after the loop we call setup_object for the first object (?),
>> but inside of the loop we skip the call for the last object (?). Am I
>> missing something, or we call ctor twice for the last object and don't
>> call it for the first one?
>
> Inside the loop we call setup_object for the "next" object. So we
> start iterating on the first one, but call setup_object for the
> second. Then the loop moves on to the second one and calls
> setup_object for the third. And so on. So the loop calls setup_object
> for every object (including the last one) except for the first one.
>
> The idea is that we want the freelist pointer that is stored in the
> current object to have a tagged pointer to the next one, so we need to
> assign a tag to the next object before storing the pointer in the
> current one.

Ah, OK, then false alarm.