This patchset aims to reduce actual memory use for guests with sparse memory. The pseries guest uses dynamic DMA windows to map the entire guest RAM but it only actually maps onlined memory which may be not be contiguous. I hit this when tried passing through NVLink2-connected GPU RAM of NVIDIA V100 and trying to map this RAM at the same offset as in the real hardware forced me to rework I handle these windows.
This moves userspace-to-host-physical translation table (iommu_table::it_userspace) from VFIO TCE IOMMU subdriver to the platform code and reuses the already existing multilevel TCE table code which we have for the hardware tables. At last in 6/6 I switch to on-demand allocation so we do not allocate huge chunks of the table if we do not have to; there is some math in 6/6. Please comment. Thanks. Alexey Kardashevskiy (6): powerpc/powernv: Remove useless wrapper powerpc/powernv: Move TCE manupulation code to its own file KVM: PPC: Make iommu_table::it_userspace big endian powerpc/powernv: Add indirect levels to it_userspace powerpc/powernv: Rework TCE level allocation powerpc/powernv/ioda: Allocate indirect TCE levels on demand arch/powerpc/platforms/powernv/Makefile | 2 +- arch/powerpc/include/asm/iommu.h | 11 +- arch/powerpc/platforms/powernv/pci.h | 44 ++- arch/powerpc/kvm/book3s_64_vio.c | 11 +- arch/powerpc/kvm/book3s_64_vio_hv.c | 18 +- arch/powerpc/platforms/powernv/pci-ioda-tce.c | 395 ++++++++++++++++++++++++++ arch/powerpc/platforms/powernv/pci-ioda.c | 192 ++----------- arch/powerpc/platforms/powernv/pci.c | 158 ----------- drivers/vfio/vfio_iommu_spapr_tce.c | 65 +---- 9 files changed, 482 insertions(+), 414 deletions(-) create mode 100644 arch/powerpc/platforms/powernv/pci-ioda-tce.c -- 2.11.0
