On Wed, 17 Apr 2013, Yasuaki Ishimatsu wrote:
Why not simply do what generic sparsemem support does by testing
PageSlab(virt_to_head_page(res)) and calling kfree() if true and freeing
back to bootmem if false? This should be like a five line patch.
Is your explanation about free_section_usemap()?
If so, I don't think we can release resource structure like
free_section_usemap().
Right, you can't release it like free_section_usemap(), but you're free to
test for PageSlab(virt_to_head_page(res)) in kernel/resource.c.
In your explanation case, memmap can be released by put_page_bootmem() in
free_map_bootmem() since all pages of memmap is used only for memmap.
But if my understanding is correct, a page of released resource structure
contain other purpose objects allocated by bootmem. So we cannot
release resource structure like free_section_usemap().
I'm thinking it would be much easier to just suppress the kfree() if
!PageSlab. If you can free an entire page with free_bootmem_late(),
that would be great,
but I'm thinking that will take more work than it's
worth. It seems fine to just do free_bootmem() and leave those pages as
reserved.
How much memory are we talking about?