Re: [Xen-devel] [PATCH 2/4] hvm/dmop: Implement copy_{to, from}_guest_buf() in terms of raw accessors

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 21.04.17 at 10:54, <andrew.cooper3@xxxxxxxxxx> wrote:
> On 21/04/2017 08:27, Jan Beulich wrote:
>>>>> On 20.04.17 at 19:59, <jennifer.herbert@xxxxxxxxxx> wrote:
>>> From: Jennifer Herbert <Jennifer.Herbert@xxxxxxxxxx>
>> Is this correct, considering that iirc the patch was new in v5 and ...
>>
>>> This also allows the usual cases to be simplified, by omitting an 
>>> unnecessary
>>> buf parameters, and because the macros can appropriately size the object.
>>>
>>> This makes copying to or from a buf that isn't big enough an error.
>>> If the buffer isnt big enough, trying to carry on regardless
>>> can only cause trouble later on.
>>>
>>> Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
>>> Signed-off-by: Jennifer Herbert <Jennifer.Herbert@xxxxxxxxxx>
>> ... this sequence of S-o-b-s?
>>
>>> --- a/xen/arch/x86/hvm/dm.c
>>> +++ b/xen/arch/x86/hvm/dm.c
>>> @@ -32,36 +32,47 @@ struct dmop_args {
>>>      struct xen_dm_op_buf buf[2];
>>>  };
>>>  
>>> -static bool copy_buf_from_guest(const xen_dm_op_buf_t bufs[],
>>> -                                unsigned int nr_bufs, void *dst,
>>> -                                unsigned int idx, size_t dst_size)
>>> +static bool _raw_copy_from_guest_buf(void *dst,
>>> +                                     const struct dmop_args *args,
>>> +                                     unsigned int buf_idx,
>>> +                                     size_t dst_bytes)
>>>  {
>>> -    size_t size;
>>> +    size_t buf_bytes;
>>>  
>>> -    if ( idx >= nr_bufs )
>>> +    if ( buf_idx >= args->nr_bufs )
>>>          return false;
>>>  
>>> -    memset(dst, 0, dst_size);
>>> +    buf_bytes =  args->buf[buf_idx].size;
>>>  
>>> -    size = min_t(size_t, dst_size, bufs[idx].size);
>>> +    if ( dst_bytes > buf_bytes )
>>> +        return false;
>> While this behavioral change is now being mentioned in the
>> description, I'm not sure I buy the argument of basically being
>> guaranteed to cause trouble down the road. Did you consider the
>> forward compatibility aspect here, allowing us to extend interface
>> structures by adding fields to their ends without breaking old
>> callers? Paul, what are your thoughts here?
> 
> DMOP is a stable ABI.  There is no legal extending of any objects.

I disagree: We have various pad fields which we check to be zero.
Putting flags there indicating presence of extended fields would be
quite fine, without breaking the ABI in any way.

> The previous semantics are guaranteed to break the ABI with future
> subops, which is why I removed it.  In the case that the guest
> originally passed an overly long buffer, and someone tried be "clever"
> here passing the same object here expecting _raw_copy_from_guest_buf()
> to DTRT, the function will end up copying too much data from the guest,
> and you will end up with something the guest wasn't intending to be part
> of the structure replacing the zero extension.

Good point, but taken care of with the flags semantics assignable
to padding fields.

> New subops which take a longer object should use a brand new object.
> 
> $MAGIC compatibility logic like you want has no business living in the
> copy helper.  Had I spotted the intention during the original dmop
> series, I would have rejected it during review.

I'm surprised you didn't notice, because iirc the behavior had
even been changed back and forth between revisions.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel