qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [RFC] create a single workqueue for each vm to update v


From: Zhanghaoyu (A)
Subject: Re: [Qemu-devel] [RFC] create a single workqueue for each vm to update vm irq routing table
Date: Thu, 28 Nov 2013 09:14:22 +0000

>>>>>      No, this would be exactly the same code that is running now:
>>>>>
>>>>>              mutex_lock(&kvm->irq_lock);
>>>>>              old = kvm->irq_routing;
>>>>>              kvm_irq_routing_update(kvm, new);
>>>>>              mutex_unlock(&kvm->irq_lock);
>>>>>
>>>>>              synchronize_rcu();
>>>>>              kfree(old);
>>>>>              return 0;
>>>>>
>>>>>      Except that the kfree would run in the call_rcu kernel thread 
>>>>> instead of
>>>>>      the vcpu thread.  But the vcpus already see the new routing table 
>>>>> after
>>>>>      the rcu_assign_pointer that is in kvm_irq_routing_update.
>>>>>
>>>>> I understood the proposal was also to eliminate the 
>>>>> synchronize_rcu(), so while new interrupts would see the new 
>>>>> routing table, interrupts already in flight could pick up the old one.
>>>> Isn't that always the case with RCU?  (See my answer above: "the 
>>>> vcpus already see the new routing table after the rcu_assign_pointer 
>>>> that is in kvm_irq_routing_update").
>>> With synchronize_rcu(), you have the additional guarantee that any 
>>> parallel accesses to the old routing table have completed.  Since we 
>>> also trigger the irq from rcu context, you know that after
>>> synchronize_rcu() you won't get any interrupts to the old destination 
>>> (see kvm_set_irq_inatomic()).
>> We do not have this guaranty for other vcpus that do not call 
>> synchronize_rcu(). They may still use outdated routing table while a 
>> vcpu or iothread that performed table update sits in synchronize_rcu().
>>
>
>Consider this guest code:
>
>   write msi entry, directing the interrupt away from this vcpu
>   nop
>   memset(&idt, 0, sizeof(idt));
>
>Currently, this code will never trigger a triple fault.  With the change to 
>call_rcu(), it may.
>
>Now it may be that the guest does not expect this to work (PCI writes are 
>posted; and interrupts can be delayed indefinitely by the pci fabric), 
>but we don't know if there's a path that guarantees the guest something that 
>we're taking away with this change.

In native environment, if a CPU's LAPIC's IRR and ISR have been pending many 
interrupts, then OS perform zeroing this CPU's IDT before receiving interrupts,
will the same problem happen?

Zhang Haoyu



reply via email to

[Prev in Thread] Current Thread [Next in Thread]