[v2,1/7] malloc: Add check for top size corruption.

  > Andreas already pointed out style issues.
>
> I'm somewhat surprised that we have accurate accounting in av->system_mem.
>
> Furthermore, for non-main arenas, I think the check should be against the
> size of a single heap, or maybe the minimum of av->system_mem and that size.

I thought about this and believe that we can ensure something more
strict: that the end of the top chunk is the same as the end of the
arena (contiguous main_arena case) or the heap (mmapped arena case),
see below. Tests passed but I'm a bit uncertain if these invariants
are always held.


Ensure that the end of the top chunk is the same as
 the end of the arena/heap.

    * malloc/malloc.c (_int_malloc): Check top size.
---
 malloc/malloc.c | 29 +++++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)
  

On 01/16/2018 01:05 PM, Istvan Kurucsai wrote:
>> Andreas already pointed out style issues.
>>
>> I'm somewhat surprised that we have accurate accounting in av->system_mem.
>>
>> Furthermore, for non-main arenas, I think the check should be against the
>> size of a single heap, or maybe the minimum of av->system_mem and that size.
> 
> I thought about this and believe that we can ensure something more
> strict: that the end of the top chunk is the same as the end of the
> arena (contiguous main_arena case) or the heap (mmapped arena case),
> see below. Tests passed but I'm a bit uncertain if these invariants
> are always held.
> 
> 
> Ensure that the end of the top chunk is the same as
>   the end of the arena/heap.
> 
>      * malloc/malloc.c (_int_malloc): Check top size.
> ---
>   malloc/malloc.c | 29 +++++++++++++++++++++++++++++
>   1 file changed, 29 insertions(+)
> 
> diff --git a/malloc/malloc.c b/malloc/malloc.c
> index f5aafd2..fd0f001 100644
> --- a/malloc/malloc.c
> +++ b/malloc/malloc.c
> @@ -2251,6 +2251,33 @@ do_check_malloc_state (mstate av)
>   }
>   #endif
> 
> +static bool
> +valid_top_chunk (mstate av, mchunkptr top)
> +{
> +  size_t size = chunksize(top);
> +
> +  assert (av);

I think we can drop that assert, it is implied by the pointer 
dereference in the subsequent assert.

> +  assert (av->top != initial_top (av));
> +
> +  if (av == &main_arena)
> +    {
> +      if ((contiguous (&main_arena)
> +          && __glibc_unlikely ((uintptr_t) top + size
> +                               != (uintptr_t) mp_.sbrk_base + av->system_mem))
> +          || (!contiguous (&main_arena)
> +              && __glibc_unlikely (size > av->system_mem)))
> +        return false;
> +    }
> +  else
> +    {
> +      heap_info *heap = heap_for_ptr (top);
> +      uintptr_t heap_end = (uintptr_t) heap + heap->size;
> +      if (__glibc_unlikely ((uintptr_t) top + size != heap_end))
> +        return false;
> +    }
> +
> +  return true;
> +}

I wonder if it is possible to write this in a slightly clearer way.

Maybe add a nested if (contiguous (&main_arena)) for the first branch, 
and directly return the value of the inner-most conditional?

Thanks,
Florian
  

On 02/20/2018 02:45 PM, Florian Weimer wrote:
> On 01/16/2018 01:05 PM, Istvan Kurucsai wrote:
>>> Andreas already pointed out style issues.
>>>
>>> I'm somewhat surprised that we have accurate accounting in 
>>> av->system_mem.
>>>
>>> Furthermore, for non-main arenas, I think the check should be against 
>>> the
>>> size of a single heap, or maybe the minimum of av->system_mem and 
>>> that size.
>>
>> I thought about this and believe that we can ensure something more
>> strict: that the end of the top chunk is the same as the end of the
>> arena (contiguous main_arena case) or the heap (mmapped arena case),
>> see below. Tests passed but I'm a bit uncertain if these invariants
>> are always held.
>>
>>
>> Ensure that the end of the top chunk is the same as
>>   the end of the arena/heap.
>>
>>      * malloc/malloc.c (_int_malloc): Check top size.
>> ---
>>   malloc/malloc.c | 29 +++++++++++++++++++++++++++++
>>   1 file changed, 29 insertions(+)
>>
>> diff --git a/malloc/malloc.c b/malloc/malloc.c
>> index f5aafd2..fd0f001 100644
>> --- a/malloc/malloc.c
>> +++ b/malloc/malloc.c
>> @@ -2251,6 +2251,33 @@ do_check_malloc_state (mstate av)
>>   }
>>   #endif
>>
>> +static bool
>> +valid_top_chunk (mstate av, mchunkptr top)
>> +{
>> +  size_t size = chunksize(top);
>> +
>> +  assert (av);
> 
> I think we can drop that assert, it is implied by the pointer 
> dereference in the subsequent assert.
> 
>> +  assert (av->top != initial_top (av));
>> +
>> +  if (av == &main_arena)
>> +    {
>> +      if ((contiguous (&main_arena)
>> +          && __glibc_unlikely ((uintptr_t) top + size
>> +                               != (uintptr_t) mp_.sbrk_base + 
>> av->system_mem))
>> +          || (!contiguous (&main_arena)
>> +              && __glibc_unlikely (size > av->system_mem)))
>> +        return false;
>> +    }
>> +  else
>> +    {
>> +      heap_info *heap = heap_for_ptr (top);
>> +      uintptr_t heap_end = (uintptr_t) heap + heap->size;
>> +      if (__glibc_unlikely ((uintptr_t) top + size != heap_end))
>> +        return false;
>> +    }
>> +
>> +  return true;
>> +}
> 
> I wonder if it is possible to write this in a slightly clearer way.
> 
> Maybe add a nested if (contiguous (&main_arena)) for the first branch, 
> and directly return the value of the inner-most conditional?

Any further comments here?

Thanks,
Florian