On 16.04.21 09:25, Oscar Salvador wrote:
On Thu, Apr 15, 2021 at 01:19:59PM +0200, David Hildenbrand wrote:
Implementation wise we will reuse vmem_altmap infrastructure to override
the default allocator used by __populate_section_memmap.
Part of the implementation also relies on memory_block structure gaining
a new field which specifies the number of vmemmap_pages at the beginning.
This patch also introduces the following functions:
- vmemmap_init_space: Initializes vmemmap pages by calling move_pfn_range_to_zone(),
calls kasan_add_zero_shadow() or the vmemmap range and marks
online as many sections as vmemmap pages fully span.
- vmemmap_adjust_pages: Accounts/substract vmemmap_pages to node and zone
present_pages
- vmemmap_deinit_space: Undoes what vmemmap_init_space does.
This is a bit asynchronous; and the function names are not really expressing what is being done :) I'll try to come up with better names below.
Yeah, was not happy either with the names but at that time I could not
come up with anything better.
It is worth mentioning that the real "mess" is that we want offline_pages() to properly handle zone->present_pages going to 0. Therefore, we want to manually mess with the present page count.
This should be explained by this:
"On offline, memory_block_offline() calls vmemmap_adjust_pages() prior to calling
offline_pages(), because offline_pages() performs the tearing-down of kthreads
and the rebuilding of the zonelists if the node/zone become empty."
Is not that clear?
Ehm, it is; for some reason my eyes ignored it -- sorry.
--
Thanks,
David / dhildenb
