Remote flushing api's does a busy wait which is fine in bare-metal scenario. But with-in the guest, the vcpus might have been pre-empted or blocked. In this scenario, the initator vcpu would end up busy-waiting for a long amount of time. This patch set implements para-virt flush tlbs making sure that it does not wait for vcpus that are sleeping. And all the sleeping vcpus flush the tlb on guest enter. Idea was discussed here: https://lkml.org/lkml/2012/2/20/157 The best result is achieved when we're overcommiting the host by running multiple vCPUs on each pCPU. In this case PV tlb flush avoids touching vCPUs which are not scheduled and avoid the wait on the main CPU. In addition, thanks for commit 9e52fc2b50d ("x86/mm: Enable RCU based page table freeing (CONFIG_HAVE_RCU_TABLE_FREE=y)") Testing on a Xeon Gold 6142 2.6GHz 2 sockets, 32 cores, 64 threads, so 64 pCPUs, and each VM is 64 vCPUs. ebizzy -M vanilla optimized boost 1VM 46799 48670 4% 2VM 23962 42691 78% 3VM 16152 37539 132% Note: The patchset is not rebased against "locking/qspinlock/x86: Avoid test-and-set when PV_DEDICATED is set" v3 since I can still observe a little improvement for 64 vCPUs on 64 pCPUs, it is due to the system is not completely isolated, there are many housekeeping tasks work sporadically, and vCPUs are preemted some times, I also confirm this when adding some print to the kvm_flush_tlb_others. After PV_DEDICATED is merged, we can disable pv tlb flush when not overcommiting if it is needed. v7 -> v8: * rebase against latest kvm/queue v6 -> v7: * don't check !flushmask * use arch_initcall() to achieve late allocate percpu mask v5 -> v6: * fix the percpu mask * rebase against latest kvm/queue v4 -> v5: * flushmask instead of cpumask v3 -> v4: * use READ_ONCE() * use try_cmpxchg instead of cmpxchg * add {} to if * no FLUSH flags to preserve during set_preempted * "KVM: X86" prefix to patch subject v2 -> v3: * percpu cpumask v1 -> v2: * a new CPUID feature bit * fix cmpxchg check * use kvm_vcpu_flush_tlb() to get the statistics right * just OR the KVM_VCPU_PREEMPTED in kvm_steal_time_set_preempted * add a new bool argument to kvm_x86_ops->tlb_flush * __cpumask_clear_cpu() instead of cpumask_clear_cpu() * not put cpumask_t on stack * rebase the patchset against "locking/qspinlock/x86: Avoid test-and-set when PV_DEDICATED is set" v3 Wanpeng Li (4): KVM: X86: Add vCPU running/preempted state KVM: X86: Add Paravirt TLB Shootdown KVM: X86: introduce invalidate_gpa argument to tlb flush KVM: X86: Add flush_on_enter before guest enter Documentation/virtual/kvm/cpuid.txt | 4 +++ arch/x86/include/asm/kvm_host.h | 2 +- arch/x86/include/uapi/asm/kvm_para.h | 5 ++++ arch/x86/kernel/kvm.c | 49 +++++++++++++++++++++++++++++++++++- arch/x86/kvm/cpuid.c | 3 ++- arch/x86/kvm/svm.c | 14 +++++------ arch/x86/kvm/vmx.c | 21 ++++++++-------- arch/x86/kvm/x86.c | 25 +++++++++++------- 8 files changed, 94 insertions(+), 29 deletions(-) -- 2.7.4
