Re: [Qemu-devel] [PATCH] macio: fix overflow in lba to offset conversion for ATAPI devices

[John Snow]

On 01/04/2016 03:54 PM, Mark Cave-Ayland wrote:
> On 04/01/16 20:36, John Snow wrote:
> 
>> On 01/04/2016 02:15 PM, Mark Cave-Ayland wrote:
>>> On 04/01/16 19:04, P J P wrote:
>>>
>>>> +-- On Mon, 4 Jan 2016, Mark Cave-Ayland wrote --+
>>>> |      /* Calculate current offset */
>>>> | -    offset = (int64_t)(s->lba << 11) + s->io_buffer_index;
>>>> | +    offset = ((int64_t)(s->lba) << 11) + s->io_buffer_index;
>>>>
>>>> Maybe ((int64_t)s->lba << 11) ? No parenthesis around s->lba.
>>>
>>> Yes that works here too (perhaps I was just being over-cautious).
>>> Alex/John, please let me know if you want me to resubmit.
>>>
>>
>> PJP's version should work just fine. I won't ask you to resubmit, though...
> 
> Great, thanks :)
> 
>> ...But, well, while we're here, I have a question for you:
>>
>> So s->lba is an int that we left shift by 11 for a max of (2^43 - 2^11)
>> then we add it against s->io_buffer_index, a uint64_t, so this statement
>> could still in theory overflow.
>>
>> Except not really, since io_buffer_index is bounded (in general) by
>> io_buffer_total_len, which is usually (IDE_DMA_BUF_SECTORS*512 + 4) ->
>> ~132K.
>>
>> I don't think there's any rigorous bounds-checking of io_buffer_index,
>> just ad-hoc checking when we're good enough to remember to do it. And we
>> don't seem to do it anywhere in macio. Is it worth peppering in an
>> assert somewhere that io_buffer_index is reasonably small?
> 
> The DBDMA engine is limited to 16-bit transfers so the maximum transfer
> size is 64K, and s->io_buffer_index is used to hold the current position
> within this transfer so unless we get some very large disks I think we
> should be okay here?
> 

For all non-malicious uses of the code, yes.

If I want to apply some rigorous checking to this bound I should just
add a function to manipulate it centrally in core.c, I think.

> 
> ATB,
> 
> Mark.
> 


I'll pull this and edit it to PJP's suggestion.

--js