Re: [PATCH] Containment measures for slab objects on scatter gather lists

[David Miller <davem@davemloft.net>]

From: Christoph Lameter <clameter@sgi.com>
Date: Thu, 28 Jun 2007 21:01:36 -0700 (PDT)

> Modify the functions in the affected arches to check for PageSlab() and 
> use a NULL mapping if such a page is encountered. This may only be 
> necessary for parisc and arm since sparc64 and xtensa do not scan over 
> processes mapping a page but I have modified those two arches also for 
> correctnesses sake since they use page_mapping() in flush_dcache_page().
> 
> Still a better solution would be to not use the slab allocator at all for 
> the objects that are used to send commands to the devices. These are not 
> permanent and grabbing a page from the pcp lists and putting it back is 
> likely as fast as performing a kmalloc.

Jens Axboe wants to get references to the page structs behind
kmalloc() allocated pages in his networking splice work.

We pass scatterlists around, but networking buffers are
composed of a kmalloc()'d data header area for packet headers
and some of the initial packet data, then a true scatterlist
of page/offset/len triplets.

Splice wants to work with pages everwhere.

This is a reocurring theme, we should provide some kind of
solution to these issues instead of wishing they would go
away.

We could make a special allocator in the networking that carves
chunks out of pages but I'm sure you'll find that about as stupid
and wasteful as I do :-)