On Tue, Nov 5, 2019 at 3:30 PM Dan Williams <dan.j.williams@xxxxxxxxx> wrote: > > On Tue, Nov 5, 2019 at 3:13 PM Sean Christopherson > <sean.j.christopherson@xxxxxxxxx> wrote: > > > > On Tue, Nov 05, 2019 at 03:02:40PM -0800, Dan Williams wrote: > > > On Tue, Nov 5, 2019 at 12:31 PM David Hildenbrand <david@xxxxxxxxxx> wrote: > > > > > The scarier code (for me) is transparent_hugepage_adjust() and > > > > > kvm_mmu_zap_collapsible_spte(), as I don't at all understand the > > > > > interaction between THP and _PAGE_DEVMAP. > > > > > > > > The x86 KVM MMU code is one of the ugliest code I know (sorry, but it > > > > had to be said :/ ). Luckily, this should be independent of the > > > > PG_reserved thingy AFAIKs. > > > > > > Both transparent_hugepage_adjust() and kvm_mmu_zap_collapsible_spte() > > > are honoring kvm_is_reserved_pfn(), so again I'm missing where the > > > page count gets mismanaged and leads to the reported hang. > > > > When mapping pages into the guest, KVM gets the page via gup(), which > > increments the page count for ZONE_DEVICE pages. But KVM puts the page > > using kvm_release_pfn_clean(), which skips put_page() if PageReserved() > > and so never puts its reference to ZONE_DEVICE pages. > > Oh, yeah, that's busted. Ugh, it's extra busted because every other gup user in the kernel tracks the pages resulting from gup and puts them (put_page()) when they are done. KVM wants to forget about whether it did a gup to get the page and optionally trigger put_page() based purely on the pfn. Outside of VFIO device assignment that needs pages pinned for DMA, why does KVM itself need to pin pages? If pages are pinned over a return to userspace that needs to be a FOLL_LONGTERM gup.
