RE: [Lse-tech] [RFC] numa slab for 2.5.41-mm1

RE: [Lse-tech] [RFC] numa slab for 2.5.41-mm1

[Luck, Tony]

Manfred Spraul wrote:

> - is it possible implement ptr_to_nodeid()
>   on all archs efficiently? It will happen for every kfree().

The best platform independent way that I came up with was to stash
the node id in the page structure ... the initial patch that Nitin
posted included code for this (and it's all my fault that this
added an extra element to the page structure).  I think that you
suggested that slab could overload the use of some existing field
if we wanted to pursue this direction.

If ptr_to_nodeid() is made a platform dependent function, then
there are some platforms that can do this very efficiently (since
the nodeid is embedded in some of the high-order address bits), and
some for which this is complex (e.g. platforms that concatenate
memory from each node).

-Tony

RE: [Lse-tech] [RFC] numa slab for 2.5.41-mm1

[Martin J. Bligh]

>> - is it possible implement ptr_to_nodeid()
>>   on all archs efficiently? It will happen for every kfree().
> 
> The best platform independent way that I came up with was to stash
> the node id in the page structure ... the initial patch that Nitin
> posted included code for this (and it's all my fault that this
> added an extra element to the page structure).  I think that you
> suggested that slab could overload the use of some existing field
> if we wanted to pursue this direction.
> 
> If ptr_to_nodeid() is made a platform dependent function, then

I think that's really the key .... the platforms should just make this
efficient, it's not something for the slab to worry about specifically.
Those of us that have virtual address space to burn can stick it in
struct page, whereas other people (eg me) might have to find some
other way to do it .... but that's the arch people's problem ;-)

> there are some platforms that can do this very efficiently (since
> the nodeid is embedded in some of the high-order address bits), and
> some for which this is complex (e.g. platforms that concatenate
> memory from each node).

M.

Re: [Lse-tech] [RFC] numa slab for 2.5.41-mm1

[David S. Miller]

   From: "Luck, Tony" <tony.luck@intel.com>
   Date: Tue, 8 Oct 2002 16:29:45 -0700 
   
   If ptr_to_nodeid() is made a platform dependent function, then
   there are some platforms that can do this very efficiently (since
   the nodeid is embedded in some of the high-order address bits), and
   some for which this is complex (e.g. platforms that concatenate
   memory from each node).

I suggest to do it like this, provide the portable version unless some
"HAVE_ARCH_PTR_TO_NODEID" CCP macro is defined in which case the
architecture can make the "address bits" optimization or similar.

The is how we usually handle such things.