Re: QPACK proposal: wrap absolute index values

[Dmitri Tikhonov <[e-mail]>]

On Mon, Aug 06, 2018 at 10:29:50AM +1000, Martin Thomson wrote:
> I've no real opinion on the wrapping proposal.  It seems like it might
> work for the Largest Reference.  The math is a tiny bit annoying to
> decode in the same way the packet number encoding is, but it is
> doable.  I imagine that you reconstruct the Largest Reference based on
> the smallest index in the table.  That is:
> 
> rotation = floor(table_size / 32)
> base = minimum - (minimum % rotation)
> # equivalently: base = floor(minimum / rotation) * rotation
> largest = encoded + base
> if largest < minimum:
>   largest += rotation

My thinking is that there is no need to reconstruct the Largest
Reference, as the encoded value represents the actual absolute
index.  It is its interpretation that requires some logic.

One way to think about it is whether the absolute index is in the
Past or the Future -- as it is always either one or the other.

        The window is 2 * MaxEntries wide

 1                                        2 * MaxEntries

 |--------------------------------------------|

If the current maximum absolute index has the value MaxEntries, all
values to its left are in the Past, while all entries to the right
are in the Future:

    <-- Past ---------\ /---- Future --->
 |----------------- CurMax -------------------|

If CurMax is larger than or equal to than MaxEntries, the Past region is
contiguous.  If CurMax is smaller than or equal to MaxEntries, the Future
region is contiguous.

The code is

    bool
    isHeaderBlocked (unsigned LargestRef)
    {
        if (CurMax <= MaxEntries)
            return LargestRef > CurMax
                                && LargestRef <= CurMax + MaxEntries;
        else
            return !(LargestRef <= CurMax
                                && LargestRef > CurMax - MaxEntries + 1);
    }


  - Dmitri.