On Wed, Jul 31, 2024, Paolo Bonzini wrote: > On Wed, Jul 31, 2024 at 1:19 PM Hao Peng <flyingpenghao@xxxxxxxxx> wrote: > > > So if anything you could check list_empty(&kvm->arch.active_mmu_pages) > > > before the loop of kvm_zap_obsolete_pages(), similar to what is done in > > > kvm_mmu_zap_oldest_mmu_pages(). I doubt it can have any practical > > > benefit, though. > > > > I did some tests, when ept=0, kvm_zap_obsolete_pages was called 42 > > times, and only 17 times > > active_mmu_page list was not empty. When tdp_mmu was enabled, > > active_mmu_page list > > was always empty. > > Did you also test with nested virtual machines running? > > In any case, we're talking of a difference of about 100 instructions > at most, so it's irrelevant. It's not even remotely close to 100 instructions. It's not even 10 instructions. It's 3 instructions, and maybe two uops? Modern compilers are smart enough to optimize usage of kvm_mmu_commit_zap_page() so that the caller inlines the list_empty(invalid_list) check, but the guts of the zap code are non-inlined. So, as is, the generated code is: 0x00000000000599a7 <+55>: mov 0x8d40(%r12),%rbp 0x00000000000599af <+63>: cmp %rbp,%r15 0x00000000000599b2 <+66>: mov 0x8(%rbp),%rbx 0x00000000000599b6 <+70>: je 0x599d6 <kvm_zap_obsolete_pages+102> 0x00000000000599d6 <+102>: mov 0x8d48(%r12),%rax 0x00000000000599de <+110>: cmp %r14,%rax 0x00000000000599e1 <+113>: je 0x59a5f <kvm_zap_obsolete_pages+239> 0x0000000000059a5f <+239>: mov 0x8(%rsp),%rax 0x0000000000059a64 <+244>: sub %gs:0x28,%rax 0x0000000000059a6d <+253>: jne 0x59a86 <kvm_zap_obsolete_pages+278> 0x0000000000059a6f <+255>: add $0x10,%rsp 0x0000000000059a73 <+259>: pop %rbx 0x0000000000059a74 <+260>: pop %rbp 0x0000000000059a75 <+261>: pop %r12 0x0000000000059a77 <+263>: pop %r13 0x0000000000059a79 <+265>: pop %r14 0x0000000000059a7b <+267>: pop %r15 0x0000000000059a7d <+269>: ret and adding an extra list_empty(kvm->arch.active_mmu_pages) generates: 0x000000000005999a <+42>: mov 0x8d38(%rdi),%rax 0x00000000000599a1 <+49>: cmp %rax,%r15 0x00000000000599a4 <+52>: je 0x59a6f <kvm_zap_obsolete_pages+255> 0x0000000000059a6f <+255>: mov 0x8(%rsp),%rax 0x0000000000059a74 <+260>: sub %gs:0x28,%rax 0x0000000000059a7d <+269>: jne 0x59a96 <kvm_zap_obsolete_pages+294> 0x0000000000059a7f <+271>: add $0x10,%rsp 0x0000000000059a83 <+275>: pop %rbx 0x0000000000059a84 <+276>: pop %rbp 0x0000000000059a85 <+277>: pop %r12 0x0000000000059a87 <+279>: pop %r13 0x0000000000059a89 <+281>: pop %r14 0x0000000000059a8b <+283>: pop %r15 0x0000000000059a8d <+285>: ret i.e. it elides the list_empty(invalid_list) check, that's it.
