Re: [PATCH] i386: Don't try to set MSR_KVM_ASYNC_PF_EN if kernel-irqchip=off

[Vitaly Kuznetsov <vkuznets@redhat.com>]

Eduardo Habkost <ehabkost@redhat.com> writes:

> On Tue, Sep 22, 2020 at 05:38:12PM +0200, Vitaly Kuznetsov wrote:
>> Eduardo Habkost <ehabkost@redhat.com> writes:
>> 
>> > This addresses the following crash when running Linux v5.8
>> > with kernel-irqchip=off:
>> >
>> >   qemu-system-x86_64: error: failed to set MSR 0x4b564d02 to 0x0
>> >   qemu-system-x86_64: ../target/i386/kvm.c:2714: kvm_buf_set_msrs: Assertion `ret == cpu->kvm_msr_buf->nmsrs' failed.
>> >
>> > There is a kernel-side fix for the issue too (kernel commit
>> > d831de177217 "KVM: x86: always allow writing '0' to
>> > MSR_KVM_ASYNC_PF_EN"), but it's nice to simply not trigger
>> > the bug if running an older kernel.
>> >
>> > Fixes: https://bugs.launchpad.net/bugs/1896263
>> > Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
>> > ---
>> >  target/i386/kvm.c | 7 ++++++-
>> >  1 file changed, 6 insertions(+), 1 deletion(-)
>> >
>> > diff --git a/target/i386/kvm.c b/target/i386/kvm.c
>> > index 9efb07e7c83..1492f41349f 100644
>> > --- a/target/i386/kvm.c
>> > +++ b/target/i386/kvm.c
>> > @@ -2818,7 +2818,12 @@ static int kvm_put_msrs(X86CPU *cpu, int level)
>> >          kvm_msr_entry_add(cpu, MSR_IA32_TSC, env->tsc);
>> >          kvm_msr_entry_add(cpu, MSR_KVM_SYSTEM_TIME, env->system_time_msr);
>> >          kvm_msr_entry_add(cpu, MSR_KVM_WALL_CLOCK, env->wall_clock_msr);
>> > -        if (env->features[FEAT_KVM] & (1 << KVM_FEATURE_ASYNC_PF)) {
>> > +        /*
>> > +         * Some kernel versions (v5.8) won't let MSR_KVM_ASYNC_PF_EN to be set
>> > +         * at all if kernel-irqchip=off, so don't try to set it in that case.
>> > +         */
>> > +        if (env->features[FEAT_KVM] & (1 << KVM_FEATURE_ASYNC_PF) &&
>> > +            kvm_irqchip_in_kernel()) {
>> >              kvm_msr_entry_add(cpu, MSR_KVM_ASYNC_PF_EN, env->async_pf_en_msr);
>> >          }
>> >          if (env->features[FEAT_KVM] & (1 << KVM_FEATURE_PV_EOI)) {
>> 
>> I'm not sure kvm_irqchip_in_kernel() was required before we switched to
>> interrupt-based APF (as we were always injecting #PF) but with
>> kernel-5.8+ this should work. [...]
>
> Were guests able to set MSR_KVM_ASYNC_PF_EN to non-zero with
> kernel-irqchip=off on hosts running Linux <= 5.7? 

lapic_in_kernel() check appeared in kernel with the following commit:

commit 9d3c447c72fb2337ca39f245c6ae89f2369de216
Author: Wanpeng Li <wanpengli@tencent.com>
Date:   Mon Jun 29 18:26:31 2020 +0800

    KVM: X86: Fix async pf caused null-ptr-deref

which was post-interrupt-based-APF. I *think* it was OK to enable APF
with !lapic_in_kernel() before (at least I don't see what would not
allow that).

> I am assuming
> kvm-asyncpf never worked with kernel-irqchip=off (and enabling it
> by default with kernel-irqchip=off was a mistake).
>
>
>>                         [...] We'll need to merge this with
>> 
>> https://lists.nongnu.org/archive/html/qemu-devel/2020-09/msg02963.html
>> (queued by Paolo) and
>> https://lists.nongnu.org/archive/html/qemu-devel/2020-09/msg06196.html
>> which fixes a bug in it.
>> 
>> as kvm_irqchip_in_kernel() should go around both KVM_FEATURE_ASYNC_PF
>> and KVM_FEATURE_ASYNC_PF_INT I believe.
>
> Shouldn't we just disallow kvm-asyncpf-int=on if kernel-irqchip=off?

(Sarcasm: if disallowing 'kernel-irqchip=off' is not an option, then)
yes, we probably can, but kvm-asyncpf-int=on is the default we have so
we can't just implicitly change it underneath or migration will break...

-- 
Vitaly