Re: [PATCH 03/12] Remove KVM_REQ_PENDING_TIMER.

[Avi Kivity <avi@redhat.com>]

On 12/02/2009 02:59 PM, Chris Lalancette wrote:
> On 12/01/2009 08:04 PM, Jan Kiszka wrote:
>    
>> Chris Lalancette wrote:
>>      
>>> KVM_REQ_PENDING_TIMER is set and cleared in a couple of places,
>>> but it never seems to be actually checked.  Remove it.
>>>
>>>        
>> I would suggest to study the introducing commit
>> 06e05645661211b9eaadaf6344c335d2e80f0ba2. My strong feeling is that this
>> removal is wrong.
>>      
> Ah, thanks.  Now I see what you mean.  In newer KVM, if you are in
> __vcpu_run() and are in the while (r>  0) loop, after you've
> done kvm_inject_pending_timer_irqs(), but before you do the next
> __vcpu_run() (and hence disabled local IRQs), you could receive an
> hrtimer callback.  In that case I guess you would hit the
> if (vcpu->requests || need_resched() || signal_pending()) test, don't
> inject the interrupt, and possibly spend a long time in the guest
> before an unrelated event causes the exit and hence causes the timer
> injection.
>
> Well, that certainly is a bit subtle :).  I think I might submit a
> patch to at least put a comment in vcpu_enter_guest() about this.
>
>    

It is, and I don't like it much either.

> Unfortunately, this does sort of put a damper on what's going on later
> in the series.  If the i8254 doesn't belong to the BSP (which it really
> shouldn't), then I don't know which vcpu to raise the bit on.  Maybe I
> can do it in the kvm_irq_delivery_to_apic() function, or
> kvm_apic_set_irq().  I'll take a look at that.
>    

We shouldn't have that bit at all.

An alternative is to make the pic/ioapic work in irq context 
(spin_lock_irq() and friends) and queue the interrupt directly from the 
hrtimer instead of the vcpu.  My worry is that we have quite a bit of 
work to do in irq context if the interrupt is broadcast and if we have 
many vcpus; maybe we can use a work_struct in that case.

-- 
error compiling committee.c: too many arguments to function