On Mon, Jul 31, 2023 at 09:00:06PM +0200, David Hildenbrand wrote: > Their logic is "if it's directly in the page table, create, hand it over. If > not, please go the slow path.". In many cases user space just touched these > pages so they are very likely in the page table. I think it has become pretty confusing, overall. In my mind 'pin_user_pages_fast()' should be functionally the same as 'pin_user_pages_unlocked()'. Places call fast if they have no idea about what is under memory, otherwise they call unlocked if you are pretty sure something is there that needs the mmap lock to resolve. If we need different behaviors a GUP flag makes the most sense. > Always honoring NUMA faults here does not sound like the improvement we > wanted to have :) ... we actually *don't* want to honor NUMA faults here. Yeah, I think that is right. We should not really use the CPU running PUP as any input to a NUMA algorithm.. If we want NUMA'ness then the PUP user should specify the affinity that makes sense. Jason
