On 14.08.19 22:56, Andrew Morton wrote: > On Wed, 14 Aug 2019 17:41:08 +0200 David Hildenbrand <[email protected]> wrote: > >> Commit a9cd410a3d29 ("mm/page_alloc.c: memory hotplug: free pages as higher >> order") assumed that any PFN we get via memory resources is aligned to >> to MAX_ORDER - 1, I am not convinced that is always true. Let's play safe, >> check the alignment and fallback to single pages. >> >> ... >> >> --- a/mm/memory_hotplug.c >> +++ b/mm/memory_hotplug.c >> @@ -646,6 +646,9 @@ static int online_pages_range(unsigned long start_pfn, >> unsigned long nr_pages, >> */ >> for (pfn = start_pfn; pfn < end_pfn; pfn += 1ul << order) { >> order = min(MAX_ORDER - 1, get_order(PFN_PHYS(end_pfn - pfn))); >> + /* __free_pages_core() wants pfns to be aligned to the order */ >> + if (unlikely(!IS_ALIGNED(pfn, 1ul << order))) >> + order = 0; >> (*online_page_callback)(pfn_to_page(pfn), order); >> } > > We aren't sure if this occurs, but if it does, we silently handle it. > > It seems a reasonable defensive thing to do, but should we add a > WARN_ON_ONCE() so that we get to find out about it? If we get such a > report then we can remove the WARN_ON_ONCE() and add an illuminating > comment. > >
Makes sense, do you want to add the WARN_ON_ONCE() or shall I resend? I was recently thinking about limiting offlining to memory blocks without holes - then also onlining would only apply to memory blocks without holes and we could simplify both paths (single zone/node, no holes) - including this check, we would always have memory block size alignments. But I am not sure yet if there is a valid use case for offlining/re-online boot memory with holes. -- Thanks, David / dhildenb
