Re: [dpdk-dev] [PATCH v1 3/4] vhost: improve device ready definition

[Maxime Coquelin <maxime.coquelin@redhat.com>]

On 6/23/20 11:02 AM, Matan Azrad wrote:
> 
> 
> From: Maxime Coquelin:
>> On 6/22/20 5:51 PM, Matan Azrad wrote:
>>>
>>>
>>> From: Maxime Coquelin:
>>>> On 6/22/20 3:43 PM, Matan Azrad wrote:
>>>>>
>>>>>
>>>>> From: Maxime Coquelin:
>>>>>> Sent: Monday, June 22, 2020 3:33 PM
>>>>>> To: Matan Azrad <matan@mellanox.com>; Xiao Wang
>>>>>> <xiao.w.wang@intel.com>
>>>>>> Cc: dev@dpdk.org
>>>>>> Subject: Re: [PATCH v1 3/4] vhost: improve device ready definition
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 6/22/20 12:06 PM, Matan Azrad wrote:
>>>>>>>
>>>>>>> Hi Maxime
>>>>>>>
>>>>>>> From: Maxime Coquelin <maxime.coquelin@redhat.com>
>>>>>>>> Sent: Monday, June 22, 2020 11:56 AM
>>>>>>>> To: Matan Azrad <matan@mellanox.com>; Xiao Wang
>>>>>>>> <xiao.w.wang@intel.com>
>>>>>>>> Cc: dev@dpdk.org
>>>>>>>> Subject: Re: [PATCH v1 3/4] vhost: improve device ready
>>>>>>>> definition
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On 6/22/20 10:41 AM, Matan Azrad wrote:
>>>>>>>>>> The issue is if you only check ready state only before and
>>>>>>>>>> after the message affecting the ring is handled, it can be
>>>>>>>>>> ready at both stages, while the rings have changed and state
>>>>>>>>>> change callback should
>>>>>>>> have been called.
>>>>>>>>> But in this version I checked twice, before message handler and
>>>>>>>>> after
>>>>>>>> message handler, so it should catch any update.
>>>>>>>>
>>>>>>>> No, this is not enough, we have to check also during some
>>>>>>>> handlers, so that the ready state is invalidated because
>>>>>>>> sometimes it will be ready before and after the message handler but
>> with different values.
>>>>>>>>
>>>>>>>> That's what I did in my example patch:
>>>>>>>> @@ -1847,15 +1892,16 @@ vhost_user_set_vring_kick(struct
>>>> virtio_net
>>>>>>>> **pdev, struct VhostUserMsg *msg,
>>>>>>>>
>>>>>>>> ...
>>>>>>>>
>>>>>>>>         if (vq->kickfd >= 0)
>>>>>>>>                 close(vq->kickfd);
>>>>>>>> +
>>>>>>>> +       vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
>>>>>>>> +
>>>>>>>> +       vhost_user_update_vring_state(dev, file.index);
>>>>>>>> +
>>>>>>>>         vq->kickfd = file.fd;
>>>>>>>>
>>>>>>>>
>>>>>>>> Without that, the ready check will return ready before and after
>>>>>>>> the kickfd changed and the driver won't be notified.
>>>>>>>
>>>>>>> The driver will be notified in the next
>>>>>>> VHOST_USER_SET_VRING_ENABLE
>>>>>> message according to v1.
>>>>>>>
>>>>>>> One of our assumption we agreed on in the design mail is that it
>>>>>>> doesn't
>>>>>> make sense that QEMU will change queue configuration without
>>>>>> enabling the queue again.
>>>>>>> Because of that we decided to force calling state callback again
>>>>>>> when
>>>>>> QEMU send VHOST_USER_SET_VRING_ENABLE(1) message even if the
>>>> queue is
>>>>>> already ready.
>>>>>>> So when driver/app see state enable->enable, it should take into
>>>>>>> account
>>>>>> that the queue configuration was probably changed.
>>>>>>>
>>>>>>> I think that this assumption is correct according to the QEMU code.
>>>>>>
>>>>>> Yes, this was our initial assumption.
>>>>>> But now looking into the details of the implementation, I find it
>>>>>> is even cleaner & clearer not to do this assumption.
>>>>>>
>>>>>>> That's why I prefer to collect all the ready checks callbacks
>>>>>>> (queue state and
>>>>>> device new\conf) to one function that will be called after the
>>>>>> message
>>>>>> handler:
>>>>>>> Pseudo:
>>>>>>>  vhost_user_update_ready_statuses() {
>>>>>>> 	switch (msg):
>>>>>>> 		case enable:
>>>>>>> 			if(enable is 1)
>>>>>>> 				force queue state =1.
>>>>>>> 		case callfd
>>>>>>> 		case kickfd
>>>>>>> 				.....
>>>>>>> 		Check queue and device ready + call callbacks if needed..
>>>>>>> 		Default
>>>>>>> 			Return;
>>>>>>> }
>>>>>>
>>>>>> I find it more natural to "invalidate" ready state where it is
>>>>>> handled (after vring_invalidate(), before setting new FD for call &
>>>>>> kick, ...)
>>>>>
>>>>> I think that if you go with this direction, if the first queue pair
>>>>> is invalidated,
>>>> you need to notify app\driver also about device ready change.
>>>>> Also it will cause 2 notifications to the driver instead of one in
>>>>> case of FD
>>>> change.
>>>>
>>>> You'll always end-up with two notifications, either Qemu has sent the
>>>> disable and so you'll have one notification for the disable and one
>>>> for the enable, or it didn't sent the disable and it will happen at
>>>> old value invalidation time and after new value is taken into account.
>>>>
>>>
>>> I don't see it in current QEMU behavior.
>>> When working MQ I see that some virtqs get configuration message while
>> they are in enabled state.
>>> Then, enable message is sent again later.
>>
>> I guess you mean the first queue pair? And it would not be in ready state as it
>> would be the initial configuration of the queue?
> 
> Even after initialization when queue is ready.
> 
>>>
>>>>> Why not to take this correct assumption and update ready state only
>>>>> in one
>>>> point in the code instead of doing it in all the configuration handlers
>> around?
>>>>> IMO, It is correct, less intrusive, simpler, clearer and cleaner.
>>>>
>>>> I just looked closer at the Vhost-user spec, and I'm no more so sure
>>>> this is a correct assumption:
>>>>
>>>> "While processing the rings (whether they are enabled or not), client
>>>> must support changing some configuration aspects on the fly."
>>>
>>> Ok, this doesn't explain how configuration is changed on the fly.
>>
>> I agree it lacks a bit of clarity.
>>
>>> As I mentioned, QEMU sends enable message always after configuration
>> message.
>>
>> Yes, but we should not do assumptions on current Qemu version when
>> possible. Better to be safe and follow the specification, it will be more robust.
>> There is also the Virtio-user PMD to take into account for example.
> 
> I understand your point here but do you really want to be ready for any configuration update in run time?
> What does it mean? How datatpath should handle configuration from control thread in run time while traffic is on?
> For example, changing queue size \ addresses must stop traffic before...
> Also changing FDs is very sensitive.
> 
> It doesn't make sense to me.
> 
> Also, according to "on the fly" direction we should not disable the queue unless enable message is coming to disable it.
> 
> In addition:
> Do you really want to toggle vDPA drivers\app for any configuration message? It may cause queue recreation for each one (at least for mlx5).

I want to have something robust and maintainable.

These messages arriving after a queue have been configured once are rare
events, but this is usually the kind of things that cause maintenance
burden.

If you look at my example patch, you will understand that with my
proposal, there won't be any more state change notification than with
your proposal when Qemu or any other Vhost-user master send a disable
request before sending the request that impact the queue state.

It just adds more robustness if this unlikely event happens, by
invalidating the ring state to not ready before doing the actual ring
configuration change. So that this config change is not missed by the
vDPA driver or the application.

Maxime