Re: [Qemu-block] [PATCH] block/parallels.c: avoid integer overflow in allocate_clusters()

[Max Reitz]

On 31.03.2017 16:54, Denis V. Lunev wrote:
> On 03/31/2017 04:47 PM, Max Reitz wrote:
>> On 31.03.2017 15:13, Peter Maydell wrote:
>>> Coverity (CID 1307776) points out that in the multiply:
>>>   space = to_allocate * s->tracks;
>>> we are trying to calculate a 64 bit result but the types
>>> of to_allocate and s->tracks mean that we actually calculate
>>> a 32 bit result. Add an explicit cast to force a 64 bit
>>> multiply.
>>>
>>> Signed-off-by: Peter Maydell <address@hidden>
>>> ---
>>> NB: compile-and-make-check tested only...
>>> ---
>>>  block/parallels.c | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/block/parallels.c b/block/parallels.c
>>> index 4173b3f..3886c30 100644
>>> --- a/block/parallels.c
>>> +++ b/block/parallels.c
>>> @@ -206,7 +206,7 @@ static int64_t allocate_clusters(BlockDriverState *bs, 
>>> int64_t sector_num,
>>>      }
>>>  
>>>      to_allocate = DIV_ROUND_UP(sector_num + *pnum, s->tracks) - idx;
>>> -    space = to_allocate * s->tracks;
>>> +    space = (int64_t)to_allocate * s->tracks;
>>>      if (s->data_end + space > bdrv_getlength(bs->file->bs) >> 
>>> BDRV_SECTOR_BITS) {
>>>          int ret;
>>>          space += s->prealloc_size;
>> I think the division is technically fine because to_allocate will
>> roughly be *pnum / s->tracks (and since *pnum is an int, the
>> multiplication cannot overflow).
>>
>> However, it's still good to fix this, but I would do it differently:
>> Make idx, to_allocate, and i all uint64_t or int64_t instead of
>> uint32_t. This would also prevent accidental overflow when storing the
>> result of the division in:
>>
>> idx = sector_num / s->tracks;
>> if (idx >= s->bat_size) {
>>     [...]
>>
>> The much greater problem to me appears to be that we don't check that
>> idx + to_allocate <= s->bat_size. I'm not sure whether there can be a
>> buffer overflow in the for loop below, but I'm not sure I really want to
>> know either... I think the block_status() call limits *pnum so that
>> there will not be an overflow, but then we should at least assert this.
>>
>> Max
>>
> technically we are protected by the check in
> 
> static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
>     BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
>     int64_t align, QEMUIOVector *qiov, int flags)
> ...
>     /* Forward the request to the BlockDriver, possibly fragmenting it */
>     total_bytes = bdrv_getlength(bs);
>     if (total_bytes < 0) {
>         ret = total_bytes;
>         goto out;
>     }
> 
>     max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align);
>     if (bytes <= max_bytes && bytes <= max_transfer) {
>         ret = bdrv_driver_preadv(bs, offset, bytes, qiov, 0);
>         goto out;
>     }
> 
> which guarantees that the request is always inside the length of the
> device. Thus we should be on the safe side with the mentioned
> access as bat_size is calculated from the size of the entire virtual
> disk.

Right, but then we wouldn't need the check on idx. With the way things
are, it looks a bit confusing. Maybe we should just make it an assertion?

assert(idx < s->bat_size && idx + to_allocate <= s->bat_size);

Max

signature.asc
Description: OpenPGP digital signature