On Tue, Jun 07, 2022 at 06:47:41PM +0000, Sean Christopherson wrote: > On Tue, Jun 07, 2022, Peter Xu wrote: > > On Tue, Jun 07, 2022 at 03:04:27PM +0000, Sean Christopherson wrote: > > > On Tue, Jun 07, 2022, Paolo Bonzini wrote: > > > > On 6/6/22 23:27, Peter Xu wrote: > > > > > On Mon, Jun 06, 2022 at 06:18:12PM +0200, Paolo Bonzini wrote: > > > > > > > However there seems to be something missing at least to me, on why it'll > > > > > > > fail a migration from 5.15 (without this patch) to 5.18 (with this patch). > > > > > > > In my test case, user_xfeatures will be 0x7 (FP|SSE|YMM) if without this > > > > > > > patch, but 0x0 if with it. > > > > > > > > > > > > What CPU model are you using for the VM? > > > > > > > > > > I didn't specify it, assuming it's qemu64 with no extra parameters. > > > > > > > > Ok, so indeed it lacks AVX and this patch can have an effect. > > > > > > > > > > For example, if the source lacks this patch but the destination has it, > > > > > > the source will transmit YMM registers, but the destination will fail to > > > > > > set them if they are not available for the selected CPU model. > > > > > > > > > > > > See the commit message: "As a bonus, it will also fail if userspace tries to > > > > > > set fpu features (with the KVM_SET_XSAVE ioctl) that are not compatible to > > > > > > the guest configuration. Such features will never be returned by > > > > > > KVM_GET_XSAVE or KVM_GET_XSAVE2." > > > > > > > > > > IIUC you meant we should have failed KVM_SET_XSAVE when they're not aligned > > > > > (probably by failing validate_user_xstate_header when checking against the > > > > > user_xfeatures on dest host). But that's probably not my case, because here > > > > > KVM_SET_XSAVE succeeded, it's just that the guest gets a double fault after > > > > > the precopy migration completes (or for postcopy when the switchover is > > > > > done). > > > > > > > > Difficult to say what's happening without seeing at least the guest code > > > > around the double fault (above you said "fail a migration" and I thought > > > > that was a different scenario than the double fault), and possibly which was > > > > the first exception that contributed to the double fault. > > > > > > Regardless of why the guest explodes in the way it does, is someone planning on > > > bisecting this (if necessary?) and sending a backport to v5.15? There's another > > > bug report that is more than likely hitting the same bug. > > > > What's the bisection you mentioned? I actually did a bisection and I also > > checked reverting Leo's change can also fix this issue. Or do you mean > > something else? > > Oooooh, sorry! I got completely turned around. You ran into a bug with the > fix. I thought that you were hitting the same issues as Mike where migrating > between hosts with different capabilities is broken in v5.15, but works in v5.18. Aha, no worry. > > > > https://lore.kernel.org/all/48353e0d-e771-8a97-21d4-c65ff3bc4192@xxxxxxxxxx > > > > That is kvm64, and I agree it could be the same problem since both qemu64 > > and kvm64 models do not have any xsave feature bit declared in cpuid 0xd, > > so potentially we could be migrating some fpu states to it even with > > user_xfeatures==0 on dest host. > > > > So today I continued the investigation, and I think what's really missing > > is qemu seems to be ignoring the user_xfeatures check for KVM_SET_XSAVE and > > continues even if it returns -EINVAL. IOW, I'm wondering whether we should > > fail properly and start to check kvm_arch_put_registers() retcode. But > > that'll be a QEMU fix, and it'll at least not causing random faults > > (e.g. double faults) in guest but we should fail the migration gracefully. > > > > Sean: a side note is that I can also easily trigger one WARN_ON_ONCE() in > > your commit 98c25ead5eda5 in kvm_arch_vcpu_ioctl_run(): > > > > WARN_ON_ONCE(kvm_lapic_hv_timer_in_use(vcpu)); > > > > It'll be great if you'd like to check that up. > > Ugh, userspace can force KVM_MP_STATE_UNINITIALIZED via KVM_SET_MP_STATE. Looks > like QEMU does that when emulating RESET. > > Logically, a full RESET of the xAPIC seems like the right thing to do. I think > we can get away with that without breaking ABI? And kvm_lapic_reset() has a > related bug where it stops the HR timer but not doesn't handle the HV timer :-/ > > diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c > index e69b83708f05..948aba894245 100644 > --- a/arch/x86/kvm/lapic.c > +++ b/arch/x86/kvm/lapic.c > @@ -2395,7 +2395,7 @@ void kvm_lapic_reset(struct kvm_vcpu *vcpu, bool init_event) > return; > > /* Stop the timer in case it's a reset to an active apic */ > - hrtimer_cancel(&apic->lapic_timer.timer); > + cancel_apic_timer(&apic->lapic_timer.timer); Needs to be: + cancel_apic_timer(apic); > > /* The xAPIC ID is set at RESET even if the APIC was already enabled. */ > if (!init_event) > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c > index 540651cd28d7..ed2c7cb1642d 100644 > --- a/arch/x86/kvm/x86.c > +++ b/arch/x86/kvm/x86.c > @@ -10912,6 +10912,9 @@ int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu, > mp_state->mp_state == KVM_MP_STATE_INIT_RECEIVED)) > goto out; > > + if (mp_state->mp_state == KVM_MP_STATE_UNINITIALIZED) > + kvm_lapic_reset(vcpu, false); > + > if (mp_state->mp_state == KVM_MP_STATE_SIPI_RECEIVED) { > vcpu->arch.mp_state = KVM_MP_STATE_INIT_RECEIVED; > set_bit(KVM_APIC_SIPI, &vcpu->arch.apic->pending_events); > The change looks reasonable, but sadly I did a quick run and it still triggers.. :-/ So there seems to be something else missing. -- Peter Xu
