On Thu, Jun 17, 2021 at 11:28:49AM +0200, David Hildenbrand wrote: > On 17.06.21 11:26, Naoya Horiguchi wrote: > > From: Naoya Horiguchi <naoya.horiguchi@xxxxxxx> > > > > Recent changes by patch "mm/page_alloc: allow high-order pages to be > > stored on the per-cpu lists" makes kernels determine whether to use pcp > > by pcp_allowed_order(), which breaks soft-offline for hugetlb pages. > > > > Soft-offline dissolves a migration source page, then removes it from > > buddy free list, so it's assumed that any subpage of the soft-offlined > > hugepage are recognized as a buddy page just after returning from > > dissolve_free_huge_page(). pcp_allowed_order() returns true for > > hugetlb, so this assumption is no longer true. > > > > So disable pcp during dissolve_free_huge_page() and > > take_page_off_buddy() to prevent soft-offlined hugepages from linking to > > pcp lists. Soft-offline should not be common events so the impact on > > performance should be minimal. And I think that the optimization of > > Mel's patch could benefit to hugetlb so zone_pcp_disable() is called > > only in hwpoison context. > > Mel, Oscar, does alloc_contig_range() now have similar issues or is it > avoided because the pageblock(s) are set MIGRATE_ISOLATE? > I'd expect MIGRATE_ISOLATE to be sufficient because they should bypass the PCP list in free_unref_page. -- Mel Gorman SUSE Labs
