Re: [PATCH/RFC] KVM: track pid for VCPU only on KVM_RUN ioctl

[Christian Borntraeger <borntraeger@de.ibm.com>]

On 19/08/14 12:31, Paolo Bonzini wrote:
> Il 19/08/2014 12:09, Christian Borntraeger ha scritto:
>>> I'm not sure, this does seem like a workaround for another
>>> limitation after all...  Gleb?
>>
>> Yes. We want to get rid of KVM_S390_INITIAL_RESET in QEMU. This comes
>> from a time, when we had another userspace prototype for KVM on s390
>> (kuli). Its really a wart that has to go. Its just that we are not
>> there yet to remove the call to KVM_S390_INITIAL_RESET. Doing so can
>> result in hard to debug errors after reboot, if an interrupt was made
>> pending just before reboot that gets delivered in the new instance.
>>
>> The new way for local interrupt read/write will probably be some
>> onereg or syncreg interface with a bitmask register and payload
>> registers. We have to solve some concurrency and implemenation issues
>> here.
> 
> Yes, I understand; the plan is fine and it's good that it was already on
> your todo list.
> 
> But since you acknowledge that KVM_S390_INITIAL_RESET will go, I'm not
> sure we want to apply this patch (except for the pid == 0 part, of
> course---that one is good).

Well, it makes todays QEMU (a lot) faster on s390 bootup with many CPUs. (According to strace on my system the first GET_FPU ioctl takes up to 0.079 sec. With 64 CPUs this sums up to several seconds.
But I understand your concern of touching generic KVM code only if really necessary. Let me know if I should send a minimal pid==0 version. (I would prefer the full version, of course).

Christian