Re: Dropping mb_map

[David Laight <david%l8s.co.uk@localhost>]

On Tue, Jan 26, 2010 at 07:28:05PM +0100, Joerg Sonnenberger wrote:
> Hi all,
> most architectures currently create a submap of the kernel map for use
> by the mbuf cluster allocator. I would like to effectively kill that and
> use the kmem_map instead. This would allow a dynamic upper limit only
> affected by the VA limit of kmem_map.
> 
> This case shouldn't be hit too often as the clusters are already
> allocated from a pool, so the locking difference shouldn't make a big
> difference. Comments?

One thought I've had (it works well in a system at work) it to count
the number of buffers allocated from each source - and to limit the
allocation that way. When rx data is queued (to a user) the allocation
would be reassigned to the user rx queue.

This is an alternative to using the target queue size for flow control.

It works better in so many ways ...

        David

-- 
David Laight: david%l8s.co.uk@localhost