Re: [PATCH V3 2/3] kvm: don't register wildcard MMIO EVENTFD on two buses

[Jason Wang <jasowang@redhat.com>]

On 09/01/2015 02:54 PM, Michael S. Tsirkin wrote:
> On Tue, Sep 01, 2015 at 12:47:36PM +0800, Jason Wang wrote:
>>
>> On 09/01/2015 12:31 PM, Michael S. Tsirkin wrote:
>>> On Tue, Sep 01, 2015 at 11:33:43AM +0800, Jason Wang wrote:
>>>> On 08/31/2015 07:33 PM, Michael S. Tsirkin wrote:
>>>>> On Mon, Aug 31, 2015 at 04:03:59PM +0800, Jason Wang wrote:
>>>>>>> On 08/31/2015 03:29 PM, Michael S. Tsirkin wrote:
>>>>>>>>>>>>>>>>> Thinking more about this, invoking the 0-length write after
>>>>>>>>>>>>>>>>>>>>>>>>> the != 0 length one would be better: it would mean we only
>>>>>>>>>>>>>>>>>>>>>>>>> handle the userspace MMIO like this.
>>>>>>>>>>>>>>>>> Right.
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>> Using current unittest. This patch is about 2.9% slower than before, and
>>>>>>>>>>>>> invoking 0-length write after is still 1.1% slower (mmio-datamatch-eventfd).
>>>>>>>>>>>>>
>>>>>>>>>>>>> /patch/result/-+%/
>>>>>>>>>>>>> /base/2957/0/
>>>>>>>>>>>>> /V3/3043/+2.9%/
>>>>>>>>>>>>> /V3+invoking != 0 length first/2990/+1.1%/
>>>>>>>>>>>>>
>>>>>>>>>>>>> So looks like the best method is not searching KVM_FAST_MMIO_BUS during
>>>>>>>>>>>>> KVM_MMIO_BUS. Instead, let userspace to register both datamatch and
>>>>>>>>>>>>> wildcard in this case. Does this sound good to you?
>>>>>>>>> No - we can't change userspace.
>>>>>>> Actually, the change was as simple as following. So I don't get the
>>>>>>> reason why.
>>>>> Because it's too late - we committed to a specific userspace ABI
>>>>> when this was merged in kernel, we must maintain it.
>>>> Ok ( Though I don't think it has real users for this now because it was
>>>> actually broken).
>>> It actually worked most of the time - you only trigger a use after free
>>> on deregister.
>>>
>> It doesn't work for amd and intel machine without ept.
> I thought it does :(
>
>>>>> Even if I thought yours is a good API (and I don't BTW - it's exposing
>>>>> internal implementation details) it's too late to change it.
>>>> I believe we should document the special treatment in kernel of zero
>>>> length mmio eventfd in api.txt? If yes, is this an exposing? If not, how
>>>> can userspace know the advantages of this and use it? For better API,
>>>> probably we need another new flag just for fast mmio and obsolete
>>>> current one by failing the assigning for zero length mmio eventfd.
>>> I sent a patch to update api.txt already as part of
>>> kvm: add KVM_CAP_IOEVENTFD_PF capability.
>>> I should probably split it out.
>>>
>>> Sorry, I don't think the api change you propose makes sense - just fix the
>>> crash in the existing one.
>>>
>> Ok, so I believe the fix should go:
>>
>> - having two ioeventfds when we want to assign zero length mmio eventfd
> You mean the in-kernel data structures?

Yes.

>
>> - change the kvm_io_bus_sort_cmp() and can handle zero length correctly
> This one's for amd/non ept, right? I'd rather we implemented the
> fast mmio optimization for these.

Agree, but we'd better fix it and backport it to stable first?