[Yes, the subject is deliberately provocative!] Having recently played with large, memory intensive, and very short lived VMs, I have realised that we spend a substantial amount of time cleaning up when a VM has terminated. It turns out that 99% of the overhead is due to the unnecessary cache cleaning that we perform when pages get unmapped from Stage-2, which only serves a single purpose: to make the page visible to a non-coherent IO subsystem when the page is being swapped out. It would make sense to drop these cache maintenance operations when userspace is either unmapping a VMA or simply exiting. Detecting the latter is easy, as KVM calls kvm_arch_flush_shadow_all(). The unmap case is harder, or at least it was until 5.2 gained the MMU notifier event, which allows subsystems to find out about the reason of an change in the page tables. And it turns out we can do even better: We can also avoid invalidating individual page mappings if we're tearing down the VM altogether, and replace it with a single TLBI VMALL, which will be much lighter on the whole system (specially if your interconnect is a bit flimsy). With the above, terminating a 64GB VM that has most of its memory mapped at S2 goes from several seconds (I've seen up to 12s) to less than a second on my D05. In general, multi-socket systems seem to benefit from this more than single socket machines (based on a non-representative sample of 4 random machines I have access to). The first patch affects most architectures, as it changes one of the core architecture hooks (in a fairly mechanical way). The rest is definitely ARM-specific. Marc Zyngier (7): KVM: Pass mmu_notifier_range down to kvm_unmap_hva_range() KVM: arm/arm64: Pass flags along Stage-2 unmapping functions KVM: arm/arm64: Condition cache maintenance on unmap with a flag KVM: arm/arm64: Condition TLB maintenance on unmap with a flag KVM: arm/arm64: Elide both CMOs and TBLIs on freeing the whole Stage-2 KVM: arm/arm64: Elide CMOs when retrying a block mapping KVM: arm/arm64: Elide CMOs when unmapping a range arch/arm/include/asm/kvm_host.h | 2 +- arch/arm64/include/asm/kvm_host.h | 2 +- arch/mips/include/asm/kvm_host.h | 2 +- arch/mips/kvm/mmu.c | 6 +- arch/powerpc/include/asm/kvm_host.h | 2 +- arch/powerpc/kvm/book3s.c | 5 +- arch/powerpc/kvm/e500_mmu_host.c | 4 +- arch/x86/include/asm/kvm_host.h | 3 +- arch/x86/kvm/mmu/mmu.c | 5 +- arch/x86/kvm/x86.c | 4 +- include/linux/kvm_host.h | 2 +- virt/kvm/arm/mmu.c | 97 +++++++++++++++++++---------- virt/kvm/kvm_main.c | 7 +-- 13 files changed, 88 insertions(+), 53 deletions(-) -- 2.20.1 _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
