On Tue, 4 Dec 2018, Michal Hocko wrote: > > This is a full revert of ac5b2c18911f ("mm: thp: relax __GFP_THISNODE for > > MADV_HUGEPAGE mappings") and a partial revert of 89c83fb539f9 ("mm, thp: > > consolidate THP gfp handling into alloc_hugepage_direct_gfpmask"). > > > > By not setting __GFP_THISNODE, applications can allocate remote hugepages > > when the local node is fragmented or low on memory when either the thp > > defrag setting is "always" or the vma has been madvised with > > MADV_HUGEPAGE. > > > > Remote access to hugepages often has much higher latency than local pages > > of the native page size. On Haswell, ac5b2c18911f was shown to have a > > 13.9% access regression after this commit for binaries that remap their > > text segment to be backed by transparent hugepages. > > > > The intent of ac5b2c18911f is to address an issue where a local node is > > low on memory or fragmented such that a hugepage cannot be allocated. In > > every scenario where this was described as a fix, there is abundant and > > unfragmented remote memory available to allocate from, even with a greater > > access latency. > > > > If remote memory is also low or fragmented, not setting __GFP_THISNODE was > > also measured on Haswell to have a 40% regression in allocation latency. > > > > Restore __GFP_THISNODE for thp allocations. > > > > Fixes: ac5b2c18911f ("mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE > > mappings") > > Fixes: 89c83fb539f9 ("mm, thp: consolidate THP gfp handling into > > alloc_hugepage_direct_gfpmask") > > At minimum do not remove the cleanup part which consolidates the gfp > hadnling to a single place. There is no real reason to have the > __GFP_THISNODE ugliness outside of alloc_hugepage_direct_gfpmask. >
The __GFP_THISNODE usage is still confined to alloc_hugepage_direct_gfpmask() for the thp fault path, we no longer set it in alloc_pages_vma() as done before the cleanup.