Re: [PATCH] Section mismatch contig_page_data and bootmem_node_data

[Johannes Weiner <hannes@saeurebad.de>]

Marcin Slusarz <marcin.slusarz@gmail.com> writes:

> On Wed, Aug 20, 2008 at 11:45:00PM +0200, Johannes Weiner wrote:
>> Hi,
>> 
>> > On Wed, 20 Aug 2008 21:05:41 +0200
>> > Marcin Slusarz <marcin.slusarz@gmail.com> wrote:
>> >
>> >> I thought about this warning today and found 2 other solutions:
>> >> 1) Mark contig_page_data as __ref (but it might hide real bugs).
>> >> 2) Remove bdata from struct pglist_data and access it directly through
>> >>    bootmem_node_data. It requires passing node number to all functions
>> >>    which use bdata, but unfortunately arch/ia64/mm/discontig.c handles
>> >>    node numbering its own way. I'm still investigating it.
>> 
>> Yeah, I gave it a shot once too but dropped it again after I looked at
>> ia64 code.
>> 
>> Perhaps we can just remove the static assignment and do it at boot up?
>
> That won't work - modpost will warn at different place about section
> mismatch. But even if it would work, we lose potentially useful
> analysis of all uses of pglist_data->bdata.

Right, but the current way of handling things completely circumvents the
section checking, no?

> But I think I found better solution - replace "struct bootmem_data *bdata"
> in struct pglist_data with "int bootmem_node;" and change all uses of bdata
> to &bootmem_node_data[struct pglist_data *->bootmem_node].

Good idea.  You don't even need a new number here, pgdat->node_id should
be usable out of the box to index into the bdata array.

> What do you think about it? Would it be acceptable?

Yes, that sounds good.

	Hannes