Making, drinking tea and reading an opus magnum from Jason Kim:
[Charset ISO-8859-1 unsupported, filtering to ASCII...]
>
>
> Michael Shalayeff wrote:
> > that's not what submaps are made for.
> > also since you would require some sort of allocator to be written
> > anyway what is the difference wrt using shmems anyway?
> >
> [ -- cut -- ]
> > also. what NEEDS a contiguos virtual space for later
> > split in smaller allocs. what is the application is
> > that you require this for?
>
> For the most part, this is an academic exercise.

you should not read kernel code

> > why not use mmap() ?
>
> Basically, I want a vm_map_entry covering an address region to be able
> to ultimately point to a vmspace (and thus, a vm_map), which then can be
> shared with another process.
>
> Like what happens now in uvmspace_share() for rfork(), except that in
> this case, I don't want to share the entire VM space, but a possibly
> large portion of it that may or may not have gaps in them (i.e. may not
> be filled up with allocated pages)
>
> Think of the application as a monitor for virtual memory, which also
> needs direct access to the memory region.
> In other words, All this has to be transparent with respect to the
> client, but ptrace() by itself is not an option (for now).
>
> The client process will be using mmap()/shmat() for other things (like
> allocating pages for use within the shared subspace) - and I want all vm
> operations in the given VM subspace (map,unmap, etc..) to be "auditable"
> as well as direct access to the contents of the memory itself.

so you have read the mmap(2) page and you know about
MAP_INHERIT and MAP_SHARED right ?
i believe that will allow you to implement what you want
w/ some additional protocol over say a socket...

> In the end, however, the original question still stands -
> Before getting out the work-clothes for an extended uvm hacking session,
> I thought I'd ask for advice/hints..
>
> So, how difficult would it be to subvert the submap mechanism for this
> purpose? If anyone has advice/hints to give, I'd appreciate it very much.

there is existing mechanism to implement what you want
over fork or exec w/o any additional kernel api.

cu

--
paranoic mickey (my employers have changed but, the name has remained)