Messages in this thread | | | Date | Tue, 12 Feb 2019 14:45:49 +0100 | From | Oscar Salvador <> | Subject | Re: [PATCH] mm,memory_hotplug: Explicitly pass the head to isolate_huge_page |
| |
On Tue, Feb 12, 2019 at 09:33:29AM +0100, Michal Hocko wrote: > > > > if (PageHuge(page)) { > > struct page *head = compound_head(page); > > - pfn = page_to_pfn(head) + (1<<compound_order(head)) - 1; > > if (compound_order(head) > PFN_SECTION_SHIFT) { > > ret = -EBUSY; > > break; > > } > > Why are we doing this, btw?
I assume you are referring to:
> > if (compound_order(head) > PFN_SECTION_SHIFT) { > > ret = -EBUSY; > > break; > > }
I thought it was in case we stumble upon a gigantic page, and commit (c8721bbbdd36 mm: memory-hotplug: enable memory hotplug to handle hugepage) confirms it.
But I am not really sure if the above condition would still hold on powerpc, I wanted to check it but it is a bit more tricky than it is in x86_64 because of the different hugetlb sizes. Could it be that the above condition is not true, but still the order of that hugetlb page goes beyond MAX_ORDER? It is something I have to check.
Anyway, I think that a safer way to check this would be using hstate_is_gigantic(), which checks whether the order of the hstate goes beyond MAX_ORDER. In the end, I think that all we care about is if we can get the pages to migrate to via the buddy allocator, since gigantic pages need to use another method.
Actually, alloc_migrate_huge_page() checks for it:
<--- static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, int nid, nodemask_t *nmask) {
if (hstate_is_gigantic(h)) return NULL; --->
Another thing is that AFAICS, as long as the memblock we try to offline contains a gigantic page, it will not be able to be offlined. Moreover, the -EBUSY we return in that case is not checked anywhere, although that is not really an issue because scan_movable_pages will skip it in the next loop.
Now, this is more rambling than anything: Maybe I am missing half of the picture, but I have been thinking for a while whether we could do better when it comes to gigantic pages vs hotplug. I think that we could try to migrate those in case any of the other nodes have a spare pre-allocated gigantic page.
-- Oscar Salvador SUSE L3
| |