On Thu, 25 Aug 2016, Michal Hocko wrote: > > I don't believe it has been an issue in the past for any archs that > > don't use thp. > > Well, fragmentation is a real problem and order-0 reclaim will be never > anywhere close to reliably provide higher order pages. Well, reclaiming > a lot of memory can increase the probability of a success but that > can quite often lead to over reclaim and long stalls. There are other > sources of high order requests than THP so this is not about THP at all > IMHO. >
Would it be possible to list the high-order allocations you are concerned about other than thp that doesn't have fallback behavior like skbuff and slub allocations? struct task_struct is an order-1 allocation and there may be order-1 slab bucket usage, but what is higher order or requires aggressive compaction to allocate? Surely you're not suggesting that order-0 reclaim cannot form order-1 memory. I am concerned about kernels that require a small memory footprint and cannot enable all of CONFIG_COMPACTION and CONFIG_MIGRATION. Embedded devices are not a negligible minority of kernels. > > , CONFIG_MIGRATION. Migration has a > > dependency of NUMA or memory hot-remove (not all popular). Compaction can > > defragment memory within single zone without reliance on NUMA. > > I am not sure I am following you here. > MIGRATION depends on (NUMA || ARCH_ENABLE_MEMORY_HOTREMOVE || COMPACTION || > CMA) && MMU > Embedded device may be UMA and not care for memory hotplug or failure handling and rely solely on order-0 and order-1 memory.
