Am 09.03.2017 um 13:33 schrieb Paolo Bonzini: > > > On 09/03/2017 13:16, Wanpeng Li wrote: >> From: Wanpeng Li <wanpeng.li@xxxxxxxxxxx> >> >> Commit b95234c84 (kvm: x86: do not use KVM_REQ_EVENT for APICv interrupt >> injection) disables interrupts before setting vcpu->mode to fix an race: >> >> | The IPI for posted interrupts may be issued between setting vcpu->mode = >> | IN_GUEST_MODE and disabling interrupts. If that happens, >> | kvm_trigger_posted_interrupt returns true, but smp_kvm_posted_intr_ipi >> | doesn't do anything about it. The guest is entered with PIR.ON, but the >> | posted interrupt IPI has not been sent and the interrupt is only delivered >> | to the guest on the next vmentry (if any). >> >> However the race has already been fixed by the side-effects of moving the >> RVI update after IN_GUEST_MODE in the commit: >> >> - The IPI for posted interrupts is issued after setting vcpu->mode = >> IN_GUEST_MODE and disabling interrupts, the posted interrupt IPI is >> delayed until the guest enters non-root mode; it is then trapped by >> the processor causing the interrupt to be injected. >> - The IPI for posted interrupts is issued between setting vcpu->mode = >> IN_GUEST_MODE and disabling interrupts, the movement of the RVI update >> after IN_GUEST_MODE in the commit will catch the new PIR, and set RVI. >> >> This patch tries to reduce the local IRQ disable time by restoring the >> place of disable local IRQ in order to improve host kernel responsibility >> to some degree. > > This should not make any difference; we're talking of a few dozen clock > cycles, and it's quite possible that the extra chance of getting a PI > interrupt in the host negates it. > > Paolo > Also, I think after all the discussion regarding handling vcpu->mode + preemption when it comes to a vcpu_kick(), keeping it that way is way cleaner. -- Thanks, David
