Re: [PATCH] CONFIG_PACKET_MMAP should depend on MMU

["Aubrey Li" <aubreylee@gmail.com>]

On 4/20/07, David Howells <dhowells@redhat.com> wrote:
> Aubrey Li <aubreylee@gmail.com> wrote:
>
> > The patch works properly on my side. But
> > 1) I'm not sure why you re-wrote alloc/free_pg_vec function, doesn't
> > the current implement work for NOMMU? I know you want to allocate the
> > entire data buffer as one contiguous lump, but is it really necessary?
>
> Yes.  It's not possible to map the whole buffer otherwise.  Think about it!
> mmap() returns _one_ reference address.  In MMU-mode, the non-contiguous
> physical buffers can be made to appear virtually contiguous by fudging the
> page tables and using the MMU.  This is not possible in NOMMU-mode.  The app
> will expect the buffer to be one contiguous lump in its address space, and
> will not be able to locate the other segments of the buffer.

Great explanation, thanks, :-)

>
> Actually, what I said is not quite true.  It is possible to map the whole
> buffer otherwise: I could lift the restriction that requires that you map the
> whole buffer or not at all, and then userspace could stitch the whole lot
> together itself.  This would then require userspace to be bimodal.
>
> > 2) So the mapped pages doesn't count into NR_FILE_MAPPED, is it a problem?
>
> Not really, no - there are no pagetables.
>
> Furthermore, issuing the PACKET_RX_RING sockopt does the entire allocation.
> Any subsequent mmaps on it have little effect.
>
> We could do that accounting though if you think it'd be better.  I don't
> suppose it hurts.
>
as checked in packet_set_ring, buffer size must be a multiple of PAGE_SIZE,
--------------------packet_set_ring------------------------
if (unlikely(req->tp_block_size & (PAGE_SIZE - 1)))

So why not use __get_free_pages rather than kmalloc, so that we have
pagetables to count?

-Aubrey