On Tue, Jun 29, 2010 at 04:07:40PM +0800, Xiao Guangrong wrote: > > > Marcelo Tosatti wrote: > > >> + > >> + if (sp->role.level > PT_PAGE_TABLE_LEVEL) > >> + return; > >> + > >> + if (sp->role.direct) > >> + return direct_pte_prefetch(vcpu, sptep); > > > > Can never happen. > > > > Marcelo, > > Thanks for your comment. You mean that we can't meet sp->role.direct here? > could you please tell me why? During my test, it can be triggered. Ah, for 1->1 emulation, right. > >> @@ -322,6 +395,7 @@ static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr, > >> user_fault, write_fault, > >> dirty, ptwrite, level, > >> gw->gfn, pfn, false, true); > >> + FNAME(pte_prefetch)(vcpu, sptep); > >> break; > >> } > > > > > > I'm afraid this can introduce regressions since it increases mmu_lock > > contention. Can you get some numbers with 4-vcpu or 8-vcpu guest and > > many threads benchmarks, such as kernbench and apachebench? (on > > non-EPT). > > > > The pte prefetch is the fast path, it only occupies little time, for the worst > case, only need read 128 byte form the guest pte, and if it prefetched success, > the #PF cause by later access will avoid, then we avoid to exit form the guest, > and walk guest pte, walk shadow pages, flush local tlb... a lots of work can be > reduced. > > Before i post this patchset firstly, i do the performance test by using unixbench, > it improved ~3.6% under EPT disable case. > (it's in the first version's chagelog) > > Today, i do the kernbench test with 4 vcpu and 1G memory, the result shows it > improved ~1.6% :-) OK, nice. > > Also prefetch should be disabled for EPT, due to lack of accessed bit. > > > > But we call mmu_set_spte() with speculative == false, it not touch the accessed bit. There is no accessed bit on EPT. So the aging code (kvm_age_rmapp) considers any present translation as accessed. There is no way to distinguish between actually accessed translations and prefetched (but unused) ones. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html