Re: [Qemu-devel] [PATCH 5/9] qapi_sized_buffer

[Stefan Berger <stefanb@linux.vnet.ibm.com>]

On 03/14/2013 05:06 PM, mdroth wrote:
> On Thu, Mar 14, 2013 at 11:24:03AM -0400, Stefan Berger wrote:
>> On 03/14/2013 11:11 AM, mdroth wrote:
>>> On Thu, Mar 14, 2013 at 10:51:49AM -0400, Stefan Berger wrote:
>>>> On 03/14/2013 10:28 AM, mdroth wrote:
>>>>> On Thu, Mar 14, 2013 at 09:39:14AM -0400, Stefan Berger wrote:
>>>>>> On 03/14/2013 08:18 AM, mdroth wrote:
>>>>>>> On Wed, Mar 13, 2013 at 09:48:11PM -0400, Stefan Berger wrote:
>>>>>>>> On 03/13/2013 07:18 PM, mdroth wrote:
>>>>>>>>> On Wed, Mar 13, 2013 at 06:00:24PM -0400, Stefan Berger wrote:
>>>>>>>>>> On 03/13/2013 04:52 PM, mdroth wrote:
>>>>>>>>>>
>>>>>>>>> Visitors don't have any knowledge of the data structures they're visiting
>>>>>>>>> outside of what we tell them via the visit_*() API.
>>>>>>>>>
>>>>>>>>> [...]
>>>>>>>>>
>>>>>>>>> For example, a visitor for a 16-element array of:
>>>>>>>>>
>>>>>>>>> typedef struct ComplexType {
>>>>>>>>>      int32_t foo;
>>>>>>>>>      char *bar;
>>>>>>>>> } ComplexType;
>>>>>>>>>
>>>>>>>>> would look something like:
>>>>>>>>>
>>>>>>>>> visit_start_carray(v, ...); // instruct visitor how to calculate offsets
>>>>>>>>> for (i = 0; i < 16; i++) {
>>>>>>>>>      visit_type_ComplexType(v, ...) // instruct visitor how to handle elem
>>>>>>>>>      visit_next_carray(v, ...); // instruct visitor to move to next offset
>>>>>>>>> }
>>>>>>>>> visit_end_carray(v, ...); // instruct visitor to finalize array
>>>>>>>> Given this example above, I think we will need the sized buffer. The
>>>>>>>> sized buffer targets  binary arrays and their encoding. If I was to
>>>>>>>> encode an 'unsigned char[n]' (e.g., n=200) using n, or n/2 or n/4
>>>>>>>> loops like above breaking it apart in u8, u16 or u32 respectively I
>>>>>>>> think this would 'not bed good' also considering the 2 bytes for tag
>>>>>>>> and length being added by ASN.1 for every such datatype
>>>>>>>> (u8,u16,u32). The sized buffer allows you to for example take a
>>>>>>>> memory page and write it out in one chunk adding a few bytes of
>>>>>>>> ASN.1 'decoration' around the actual data.
>>>>>>> You could do it with this interface as well actually. The Visitor will
>>>>>>> need to maintain some internal state to differentiate what it does with
>>>>>>> subsequent visit_type*/visit_next_carray/visit_end_carry. There's no
>>>>>>> reason it couldn't also track the elem size so it could tag a buffer
>>>>>>> "en masse" when visit_end_carray() gets called.
>>>>>> It depends on what you pass into visit_start_carray. In your case if
>>>>>> you pass in ComplexType you would pass in a sizeof(ComplexType) for
>>>>>> the size of each element presumably. The problem is now you havechar *foo, a string pointer, hanging off of this structure. How
>>>>>> would you handle that? Serializing ComplexType's foo and pointer
>>>>>> obviously won't do it.
>>>>> Why not?  visit_type_ComplexType() knows how to deal with
>>>>> the individual fields, including the string pointer. I'm not sure
>>>>> what's at issue here.
>>>>>
>>>>> In this case the handling for ComplexType would look something like:
>>>>>
>>>>> visit_type_Complex:
>>>>>      visit_start_struct
>>>>>      visit_type_uin32 //foo
>>>>>      visit_type_str //bar
>>>>>      visit_end_struct
>>>>>
>>>>> Granted, strings are easier to deal with. If char * was instead a plain
>>>>> old uint8_t*, we'd need a nested call to start_carray for each element.
>>>>> in this case it would look something like:
>>>>>
>>>>> visit_type_Complex:
>>>>>      visit_start_struct
>>>>>      visit_type_uin32 //foo field
>>>>>      visit_start_carray //bar field
>>>>>      for (i = 0; i < len_of_bar; i++):
>>>>>          visit_type_uint8
>>>>>          visit_next_carray
>>>>>      visit_end_carray
>>>> You really want to create a separate element for each element in
>>>> this potentially large binary array? I guess it depends on the
>>>> underlying data, but this has the potential of generating a lot of
>>>> control code around each such byte... As said, for ASN.1 encoding,
>>>> each such byte would be decorated with a tag and a length value,
>>>> consuming 2 more bytes per byte.
>>> I addressed this earlier. Your visitor doesn't have tag each
>>> element: if it know it's handling an array (because we told it via
>>> start_carray()), it can buffer them internally and tag the array en
>>> masse when end_carray() is issued.
>> If we were to do this using carray on an array of structs of the
>> following type
>>
>> struct SimpleStruct {
>>      uint8_t a;
>>      uint8_t b;
>>      uint32_t c;
>> }
>>
>> then the serialization of a and b would be buffered and flushed once
>> the 32bit output visitor (or any other than uint8_t output visitor)
>> would be called?
> I don't quite understand. For a struct, we'd tag each field
> individually, right?
>
> It's avoiding the need to tag each element in a list, each SimpleStruct,
> that's at issue, right? We have a special case for u8 arrays that is
> currently handled by qapi_sized_buffer(), and now we're trying to
> generalize this optimized handling for more complex data types via
> visit_carray_*?

I have to admit there's one mistake in the sized buffer implementation 
and that is it should take the width of each element so each element can 
be encoded in network byte order, which I think is the key point here so 
the bytestream becomes portable Using the element width we can then walk 
the array of n elements and write them out in network byte order in one 
go. I also see the sized buffer more as a generalization of the string 
visitor that just happens to work on a null terminated array of bytes.

Using the carray it seems we would create special cases in the 
implementation for when a u8, u16, u32, or u64 follows versus a string, 
list or any more complex data type. The examples above seem to need to 
buffer the u8 and then write them out versus when it was to handle an 
array of structs. With the sized buffer we can do this all with a single 
call.

> I guess my first question is whether or not it's possible for more complex
> data types. For u8 arrays we seem to use a special OCTET_STRING encoding.
>
> If we're not sure we can do this more generally, we do have the option
> of only special-casing u8 arrays to use the OCTET_STRING encoding, and
> handle the others with "non-optimized" encodings.
>
> If you have an idea for what a generalized, optimized encoding that's
> applicable for non-u8 types would look like, we can work through that if
> you have an example optimized encoding for, say, and array of
> SimpleStruct.
For an array of SimpleStruct we use the carray. No endianess conversion 
is necessary here for the structure as a whole but each u16,u32, u64 
inside such a structure will be written out correctly with their 
respective visitor. A u16[] or u32[] inside that structure would then be 
handled with a sized buffer walking the array and normalizing each 
element's endianess.

Strictly speaking, for ASN.1 encoding we don't need a carray visitor 
(but maybe other visitor types need it). It can be simulated with the 
struct visitor, which in effect also causes one more level of nesting, 
just that the identifier would be different, which really is the only 
difference then.

    Stefan