Re: [PATCH 4/7] ppc64 - Specify amount of kernel memory at boot time

[Mel Gorman <mel@csn.ul.ie>]

On Thu, 23 Feb 2006, Dave Hansen wrote:

> On Thu, 2006-02-23 at 17:19 +0000, Mel Gorman wrote:
> > On Thu, 23 Feb 2006, Dave Hansen wrote:
> > > > +/* Initialise the size of each zone in a node */
> > > > +void __init zone_sizes_init(unsigned int nid,
> > > > +		unsigned long kernelcore_pages,
> > > > +		unsigned long *zones_size)
> > >
> > > Minor nit territory: set_zone_sizes(), maybe?
> >
> > In this case, the choice of name is to match an x86 function that does
> > something very similar. If one had read through the x86 code and then saw
> > this function, it would set their expectations of what the code is
> > intended to do.
>
> x86 is bad.  Try to do better. :)

set_zone_sizes() it is.

> > per-node. A node goes no ZONE_EASYRCLM pages if it is not large enough to
> > contain kernelcore_pages. That means that on a system with 2 nodes,
> > kernelcore=512MB will results in 1024MB of ZONE_DMA in total.
> >
> > > Also, how do we want to distribute kernelcore memory over each node?
> > > The way it is coded up for now, it will all be sliced out of the first
> > > node.  I'm not sure that's a good thing.
> >
> > It gets set in every node.
>
> They hypervisor has a memory allocator which it uses to piece out memory
> to various LPARs.  When things like partition memory resizing or reboots
> occur, that memory gets allocated and freed and so forth.
>
> These machines are _also_ NUMA.  Memory can effectively get lumped so
> that you can't always get memory on a single NUMA node.  Because of all
> of the other actions of other partitions, these conditions are always
> changing.  The end result is that the amount of memory which each NUMA
> node has *in* *a* *single* *partition* can theoretically change between
> reboots.

wow. Ok, I didn't know that. It totally rules out any future option where 
an admin can say how much kernelcore they want in each node. They won't 
know for sure the size of the node in advance or the number of nodes for 
that matter.

> OK, back to the hapless system admin using kernelcore. They have a
> 4-node system with 2GB of RAM in each node for 8GB total.  They use
> kernelcore=1GB.  They end up with 4x1GB ZONE_DMA and 4x1GB
> ZONE_EASYRCLM.  Perfect.  You can safely remove 4GB of RAM.
>
> Now, imagine that the machine has been heavily used for a while, there
> is only 1 node's memory available, but CPUs are available in the same
> places as before.  So, you start up your partition again have 8GB of
> memory in one node.  Same kernelcore=1GB option.  You get 1x7GB ZONE_DMA
> and 1x1GB ZONE_EASYRCLM.  I'd argue this is going to be a bit of a
> surprise to the poor admin.

That sort of surprise is totally unacceptable but the behaviour of 
kernelcore needs to be consistent on both the x86 and the ppc (any any 
other ar. How about;

1. kernelcore=X determines the total amount of memory for !ZONE_EASYRCLM
   (be it ZONE_DMA, ZONE_NORMAL or ZONE_HIGHMEM)
2. For every node that can have ZONE_EASYRCLM, split the kernelcore across
   the nodes as a percentage of the node size

   Example: 4 nodes, 1 GiB each, kernelcore=512MB
		node 0 ZONE_DMA = 128MB
		node 1 ZONE_DMA = 128MB
		node 2 ZONE_DMA = 128MB
		node 3 ZONE_DMA = 128MB

	    2 nodes, 3GiB and 1GIB, kernelcore=512MB
		node 0 ZONE_DMA = 384
		node 1 ZONE_DMA = 128

It gets a bit more complex on NUMA for x86 because ZONE_NORMAL is 
involved but the idea would essentially be the same.

> > zones_size[] is what free_area_init() expects to receive so there is not a
> > lot of room to fiddle with it's meaning without causing more trouble.
>
> It is just passed in there as an argument.  If you can think of a way to
> make it more understandable, change it in your architecture, and send
> the patch for the main one.

I'll think about it when my head finishes crunching through the sizing of 
kernelcore.

> > > One other thing, I want to _know_ that variables being compared are in
> > > the same units.  When one is called "pages_" something and the other is
> > > something "_size", I don't _know_.
> >
> > chunk_num_pages ?
>
> No. :)  The words "chunks" and "clumps" have a bit of a stigma, just
> like the number "3".  Ask Matt Dobson.

It sounds like a touchy subject :)

> num_pages_WHAT?  node_num_pages?  silly_num_pages?
>
> Chunk is pretty meaningless.

thingie, bit, bob, piece? :)

I'll think of something.

> > yep. get_zholes_size() could be split into two functions
> > find_start_easyrclm_pfn() and get_nid_zholes_size(). Would that be pretty
> > clear-cut?
>
> I think so.

--
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/