On 04.12.20 18:50, Pavel Tatashin wrote: >>> Yes, this indeed could be a problem for some configurations. I will >>> add your comment to the commit log of one of the patches. >> >> It sounds like there is some inherent tension here, breaking THP's >> when doing pin_user_pages() is a really nasty thing to do. DMA >> benefits greatly from THP. >> >> I know nothing about ZONE_MOVABLE, is this auto-setup or an admin >> option? If the result of this patch is standard systems can no longer >> pin > 80% of their memory I have some regression concerns.. > > ZONE_MOVABLE can be configured via kernel parameter, or when memory > nodes are onlined after hot-add; so this is something that admins > configure. ZONE_MOVABLE is designed to gurantee memory hot-plug > functionality, and not availability of THP, however, I did not know > about the use case where some admins might configure ZONE_MOVABLE to > increase availability of THP because pages are always migratable in > them. The thing is, if we fragment ZONE_MOVABLE by pinning pages in > it, the availability of THP also suffers. We can migrate pages in > ZONE_NORMAL, just not guaranteed, so we can create THP in ZONE_NORMAL > as well, which is the usual case.
Right, we should document this at some place to make admins aware of this. Something like "Techniques that rely on long-term pinnings of memory (especially, RDMA and vfio) are fundamentally problematic with ZONE_MOVABLE and, therefore, memory hotunplug. Pinned pages cannot reside on ZONE_MOVABLE, to guarantee that memory can still get hotunplugged - be aware that pinning can fail even if there is plenty of free memory in ZONE_MOVABLE. In addition, using ZONE_MOVABLE might make page pinning more expensive, because pages have to be migrated off that zone first." BTW, you might also want to update the comment for ZONE_MOVABLE in include/linux/mmzone.h at the end of this series, removing the special case of pinned pages (1.) and maybe adding what happens when trying to pin pages on ZONE_MOVABLE. -- Thanks, David / dhildenb
