Re: A comment about changing kernels

[Espen Skoglund]

[Jonathan S Shapiro]
> On Mon, 2005-10-31 at 01:05 +0100, Marcus Brinkmann wrote:
>> At Fri, 28 Oct 2005 20:34:09 +0200,
>> Bernhard Kauer <address@hidden> wrote:
>>> Now to copy(): I know no functional argument to introduce a copy()
>>> into L4.sec. The only argument is performance. Because mapping (or
>>> copy) a return endpoint with every RPC will be too expensive,
>> 
>> Ask Espen about "map-once" mappings to learn how to allow to
>> optimize reply capabilities.

> No no. Map once mappings are another thing entirely. They decidedly
> do NOT provide any optimization of reply capabilities. What they
> *do* accomplish is to utterly break the notion of capability
> transfer in the same way that the mach reply port does.

I believe what Marcus is referring to here is a private conversation I
had to him about something I called "unique mappings" (feel free to
come up with a better terms).

These unique mappings work as regular mappings with the following
exceptions:

  o There can exist only one unique mapping *from* a given resource
    (page frame, endpoint, etc.) within an address space.

  o If the resource is later "mapped unique" again the old mapping is
    implicitly revoked.

Apart from these differences unique mappings work identical to regular
mappings.  What unique mappings give you is the ability to implement
lazy mapping database updates.  That is, in the majority of cases
where unique mappings are used the mapping database need not be
updated.

        eSk