On Fri, Mar 25, 2022, Paolo Bonzini wrote: > On 3/21/22 23:43, Ben Gardon wrote: > > Currently disabling dirty logging with the TDP MMU is extremely slow. > > On a 96 vCPU / 96G VM it takes ~256 seconds to disable dirty logging > > with the TDP MMU, as opposed to ~4 seconds with the legacy MMU. This > > series optimizes TLB flushes and introduces in-place large page > > promotion, to bring the disable dirty log time down to ~3 seconds. > > > > Testing: > > Ran KVM selftests and kvm-unit-tests on an Intel Haswell. This > > series introduced no new failures. > > Thanks, looks good. The one change I'd make is to just place the outcome of > build_tdp_shadow_zero_bits_mask() in a global (say tdp_shadow_zero_check) at > kvm_configure_mmu() time. The tdp_max_root_level works as a conservative > choice for the second argument of build_tdp_shadow_zero_bits_mask(). > > No need to do anything though, I'll handle this later in 5.19 time (and > first merge my changes that factor out the constant part of > vcpu->arch.root_mmu initialization, since this is part of the same ideas). This fell through the cracks. Ben is on a long vacation, I'll find my copy of the Necronomicon and do a bit of resurrection, and address the feedback from v2 along the way.
