On Thu, Feb 18, 2021 at 12:27:58PM -0500, Zi Yan wrote: > On 18 Feb 2021, at 12:25, Jason Gunthorpe wrote: > > > On Thu, Feb 18, 2021 at 02:45:54PM +0000, Matthew Wilcox wrote: > >> On Wed, Feb 17, 2021 at 11:02:52AM -0800, Andrew Morton wrote: > >>> On Wed, 17 Feb 2021 10:49:25 -0800 Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote: > >>>> page structs are not guaranteed to be contiguous for gigantic pages. The > >>> > >>> June 2014. That's a long lurk time for a bug. I wonder if some later > >>> commit revealed it. > >> > >> I would suggest that gigantic pages have not seen much use. Certainly > >> performance with Intel CPUs on benchmarks that I've been involved with > >> showed lower performance with 1GB pages than with 2MB pages until quite > >> recently. > > > > I suggested in another thread that maybe it is time to consider > > dropping this "feature" > > You mean dropping gigantic page support in hugetlb? No, I mean dropping support for arches that want to do: tail_page != head_page + tail_page_nr because they can't allocate the required page array either virtually or physically contiguously. It seems like quite a burden on the core mm for a very niche, and maybe even non-existant, case. It was originally done for PPC, can these PPC systems use VMEMMAP now? > > The cost to fix GUP to be compatible with this will hurt normal > > GUP performance - and again, that nobody has hit this bug in GUP > > further suggests the feature isn't used.. > > A easy fix might be to make gigantic hugetlb page depends on > CONFIG_SPARSEMEM_VMEMMAP, which guarantee all struct pages are contiguous. Yes, exactly. Jason
