Messages in this thread | | | Date | Tue, 24 Aug 2010 12:38:50 +0300 | From | Avi Kivity <> | Subject | Re: [PATCH v5 09/12] Retry fault before vmentry |
| |
On 08/24/2010 12:33 PM, Gleb Natapov wrote: > > >>> @@ -505,6 +506,37 @@ out_unlock: >>> return 0; >>> } >>> >>> +static int FNAME(page_fault_other_cr3)(struct kvm_vcpu *vcpu, gpa_t cr3, >>> + gva_t addr, u32 error_code) >>> +{ >>> + int r = 0; >>> + gpa_t curr_cr3 = vcpu->arch.cr3; >>> + >>> + if (curr_cr3 != cr3) { >>> + /* >>> + * We do page fault on behalf of a process that is sleeping >>> + * because of async PF. PV guest takes reference to mm that cr3 >>> + * belongs too, so it has to be valid here. >>> + */ >>> + kvm_set_cr3(vcpu, cr3); >>> + if (kvm_mmu_reload(vcpu)) >>> + goto switch_cr3; >>> + } >> With nested virtualization, we need to switch cr0, cr4, and efer as well... >> > On SVM or VMX or both?
Both. Let's defer this patch since it's an optimization, this is really complicated.
>>> + >>> + r = FNAME(page_fault)(vcpu, addr, error_code, true); >>> + >>> + if (kvm_check_request(KVM_REQ_MMU_SYNC, vcpu)) >>> + kvm_mmu_sync_roots(vcpu); >> Why is this needed? >> > http://www.mail-archive.com/kvm@vger.kernel.org/msg37827.html > > KVM_REQ_MMU_SYNC request generated here must be processed before > switching to a different cr3 (otherwise vcpu_enter_guest will process it > with the wrong cr3 in place).
Ah, it should be part of the cr3 switch block above.
>>> + >>> +switch_cr3: >>> + if (curr_cr3 != vcpu->arch.cr3) { >>> + kvm_set_cr3(vcpu, curr_cr3); >>> + kvm_mmu_reload(vcpu); >>> + } >>> + >>> + return r; >>> +} >> This has the nasty effect of flushing the TLB on AMD. >> > What is more expansive reenter the guest and handle one more fault, or > flash TLB here?
No idea. Probably the reentry. On Intel the tlb is flushed anyway.
-- error compiling committee.c: too many arguments to function
| |