Re: [dpdk-dev] [PATCH] drivers/net: update link status

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 4/18/2018 1:08 PM, Tiwei Bie wrote:
> On Wed, Apr 18, 2018 at 12:44:54PM +0100, Ferruh Yigit wrote:
>> On 4/18/2018 12:36 PM, Tiwei Bie wrote:
>>> On Wed, Apr 18, 2018 at 11:42:56AM +0100, Ferruh Yigit wrote:
>>>> On 4/18/2018 7:49 AM, Tan, Jianfeng wrote:
>>>>> Hi Ferruh,
>>>>>
>>>>>
>>>>> On 4/17/2018 7:26 PM, Ferruh Yigit wrote:
>>>>>> On 4/17/2018 5:54 AM, Tiwei Bie wrote:
>>>>>>> On Mon, Apr 16, 2018 at 05:10:24PM +0100, Ferruh Yigit wrote:
>>>>>>>> On 4/14/2018 11:55 AM, Tiwei Bie wrote:
>>>>>>>>> On Fri, Apr 13, 2018 at 10:53:55PM +0100, Ferruh Yigit wrote:
>>>>>>>>>> On 4/10/2018 4:41 PM, Tiwei Bie wrote:
>>>>>>>>>>> On Tue, Mar 13, 2018 at 06:05:34PM +0000, Ferruh Yigit wrote:
>>>>>>>>>>>> Update link status related feature document items and minor updates in
>>>>>>>>>>>> some link status related functions.
>>>>>>>>>>>>
>>>>>>>>>>>> Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
>>>>>>>>>>>> ---
>>>>>>>>>>>>   doc/guides/nics/features/fm10k.ini      | 2 ++
>>>>>>>>>>>>   doc/guides/nics/features/fm10k_vf.ini   | 2 ++
>>>>>>>>>>>>   doc/guides/nics/features/i40e_vf.ini    | 1 +
>>>>>>>>>>>>   doc/guides/nics/features/igb_vf.ini     | 1 +
>>>>>>>>>>>>   doc/guides/nics/features/qede.ini       | 1 -
>>>>>>>>>>>>   doc/guides/nics/features/qede_vf.ini    | 1 -
>>>>>>>>>>>>   doc/guides/nics/features/vhost.ini      | 2 --
>>>>>>>>>>>>   doc/guides/nics/features/virtio_vec.ini | 1 +
>>>>>>>>>>>>   drivers/net/e1000/em_ethdev.c           | 2 +-
>>>>>>>>>>>>   drivers/net/ena/ena_ethdev.c            | 2 +-
>>>>>>>>>>>>   drivers/net/fm10k/fm10k_ethdev.c        | 6 ++----
>>>>>>>>>>>>   drivers/net/i40e/i40e_ethdev_vf.c       | 2 +-
>>>>>>>>>>>>   drivers/net/ixgbe/ixgbe_ethdev.c        | 2 +-
>>>>>>>>>>>>   drivers/net/mlx4/mlx4_ethdev.c          | 2 +-
>>>>>>>>>>>>   drivers/net/mlx5/mlx5_ethdev.c          | 2 +-
>>>>>>>>>>>>   15 files changed, 15 insertions(+), 14 deletions(-)
>>>>>>>>>>> [...]
>>>>>>>>>>>> diff --git a/doc/guides/nics/features/vhost.ini b/doc/guides/nics/features/vhost.ini
>>>>>>>>>>>> index dffd1f493..31302745a 100644
>>>>>>>>>>>> --- a/doc/guides/nics/features/vhost.ini
>>>>>>>>>>>> +++ b/doc/guides/nics/features/vhost.ini
>>>>>>>>>>>> @@ -4,8 +4,6 @@
>>>>>>>>>>>>   ; Refer to default.ini for the full list of available PMD features.
>>>>>>>>>>>>   ;
>>>>>>>>>>>>   [Features]
>>>>>>>>>>>> -Link status          = Y
>>>>>>>>>>>> -Link status event    = Y
>>>>>>>>>>> I think vhost PMD supports above features.
>>>>>>>>>> I am not able to find where it is supported.
>>>>>>>>>>
>>>>>>>>>> Some virtual PMDs report fixed link, with empty link_update() dev_ops, and they
>>>>>>>>>> are not reported as supporting Link status, as far as I can see vhost also one
>>>>>>>>>> of them.
>>>>>>>>>>
>>>>>>>>>> And for Link status event, PMD needs to support LSC interrupts and should
>>>>>>>>>> register interrupt handler for it, which I can't find for vhost.
>>>>>>>>>>
>>>>>>>>>> I will send next version without updating above one, please point me where these
>>>>>>>>>> support added if I missed them.
>>>>>>>>> In drivers/net/vhost/rte_eth_vhost.c you could find below functions:
>>>>>>>>>
>>>>>>>>> static int
>>>>>>>>> new_device(int vid)
>>>>>>>>> {
>>>>>>>>> 	......
>>>>>>>>>
>>>>>>>>> 	eth_dev->data->dev_link.link_status = ETH_LINK_UP;
>>>>>>>>>
>>>>>>>>> 	......
>>>>>>>>>
>>>>>>>>> 	_rte_eth_dev_callback_process(eth_dev, RTE_ETH_EVENT_INTR_LSC, NULL);
>>>>>>>>>
>>>>>>>>> 	......
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> static void
>>>>>>>>> destroy_device(int vid)
>>>>>>>>> {
>>>>>>>>> 	......
>>>>>>>>>
>>>>>>>>> 	eth_dev->data->dev_link.link_status = ETH_LINK_DOWN;
>>>>>>>>>
>>>>>>>>> 	......
>>>>>>>>>
>>>>>>>>> 	_rte_eth_dev_callback_process(eth_dev, RTE_ETH_EVENT_INTR_LSC, NULL);
>>>>>>>>>
>>>>>>>>> 	......
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> They are the callbacks for vhost library.
>>>>>>>>>
>>>>>>>>> When a frontend (e.g. QEMU) is connected to this vhost backend
>>>>>>>>> and the frontend virtio device becomes ready, new_device() will
>>>>>>>>> be called by the vhost library, and the link status will be
>>>>>>>>> updated to UP.
>>>>>>>>>
>>>>>>>>> And when e.g. the connection is closed, destroy_device() will be
>>>>>>>>> called by the vhost library, and the link status will be updated
>>>>>>>>> to DOWN.
>>>>>>>>
>>>>>>>> Got it. This behavior is similar for virtual PMDs. Provide static link
>>>>>>>> information and update link as UP during start and update it as DOWN during stop.
>>>>>>> No, the link status isn't updated during vhost PMD start
>>>>>>> and stop. When the vhost PMD has been started, the link
>>>>>>> status still may be DOWN. The link status becomes UP only
>>>>>>> when the QEMU (it's another virtual machine process which
>>>>>>> has a virtio device) connects to this vhost PMD via a UNIX
>>>>>>> socket and the virtio driver in the virtual machine has
>>>>>>> setup the virtio device of the virtual machine.
>>>>>>>
>>>>>>> So if vhost PMD reports the link status as DOWN, it means
>>>>>>> there is no QEMU (virtual machine) connects to it or the
>>>>>>> virtio device in the virtual machine hasn't been setup.
>>>>>>> (PS. The frontend can also be virtio-user PMD besides QEMU)
>>>>>> I believe announcing link feature reporting on virtual pmds still in gray area,
>>>>>> but because of qemu involvement in vhost case, I will keep link feature but will
>>>>>> drop link event.
>>>>>
>>>>> AFAIK, link status means we can get link status through APIs like 
>>>>> rte_eth_link_get(); while link status event means applications can 
>>>>> register link status events, and those events get called if link status 
>>>>> is changed.
>>>>>
>>>>> If I understand it correctly, for vhost, we can keep both link status 
>>>>> and link status event for vhost.
>>>>>
>>>>> Could you specify the reason why we remove link status event?
>>>>
>>>> Hi Jianfeng,
>>>>
>>>> I think problem is the definition of the features are not clear, that is why I
>>>> started a doc to document them (doc/guides/nics/features.rst)
>>>>
>>>> "Link status", I think we agree on this one. PMD should provide up-to-date,
>>>> valid link data on rte_eth_dev_data->dev_link. So that this link information can
>>>> be get by rte_eth_link_get(), rte_eth_link_get_nowait() ethdev APIs.
>>>>
>>>> rte_eth_link_get() uses dev_ops->link_update() to get the latest link
>>>> information, for virtual PMDs, including vhost, this function does nothing,
>>>> because there is no actual physical link. That is why I was not sure advertising
>>>> link status feature for vhost, and other virtual PMDs doesn't report this
>>>> feature. After Tiwei's comment that link status shows that qemu connected to
>>>> vhost, added this feature back to vhost PMD.
>>>>
>>>>
>>>> "Link status event", can be
>>>> a- PMD calls application callback in link change
>>>> b- PMD registers interrupt handler for link status change interrupts
>>>>
>>>> Based on how other PMDs report this feature, I believe it is (b), and I have
>>>> documented that way. And vhost "Link status event" feature removed based on this.
>>>>
>>>> There are some set of config options and flags to control the LSC interrupt,
>>>> that also effects rte_eth_link_get() and rte_eth_link_get_nowait() APIs, I
>>>> believe that is the main concern here.
>>>>
>>>> As commented above, I don't understand why calling user register callback for
>>>> link change event is something on PMD decision, it should be default behavior
>>>> for PMD.
>>>
>>> Do you mean you don't understand why vhost PMD calls
>>> _rte_eth_dev_callback_process(eth_dev, RTE_ETH_EVENT_INTR_LSC, NULL)
>>> in new_device() and destroy_device()? While other physical
>>> NIC drivers call this in their interrupt handlers which
>>> are registered via rte_intr_callback_register().
>>
>> No, I understand this part.
>>
>> My question is why not all PMDs call callback_process() when link changed as a
>> requirement? Why PMD is giving decision to call or not call callback_process() ?
> 
> I got your point now. Maybe because in the library level,
> it just knows an interrupt happened but doesn't know whether
> it's a LSC interrupt or not (sometimes it needs to check
> some vendor specific registers). So the simplest way is
> just to ask each PMD to call callback_process() directly.

Yep.

> But unfortunately, some PMDs didn't do it as we expected..
> (I didn't look into other PMDs, so I'm not sure about the
> details.)

Yes, some PMDs doesn't call callback_process() even they support link status and
link status interrupts. This may be because this is not documented clearly.
Best option can be handling this in ethdev somehow.

> 
> Thanks
> 
> 
>>
>>>
>>> The new_device() and destroy_device() in vhost PMD can be
>>> treated as something like the interrupt handlers in physical
>>> NIC drivers. They are the callbacks which are registered via
>>> rte_vhost_driver_callback_register() and will be called by
>>> vhost library when vhost events happen (and they should be
>>> translated to link status change events when we try to wrap
>>> it as a net PMD).
>>>
>>> Thanks
>>>
>>>> What is the point of leaving this into PMD and think this as a feature
>>>> of PMD?
>>>>
>>>>
>>>> Overall, practical reason of this table is to inform developer/user about PMD
>>>> features, which is indeed device + driver features, and help her to in
>>>> development or on setting expectations. We can always discuss what helps more to
>>>> developer/user and update the features table.
>>>>
>>>> Thanks,
>>>> ferruh
>>>>
>>>>>
>>>>> Thanks,
>>>>> Jianfeng
>>>>>
>>>>>>
>>>>>> Will send a new patch to reflect this.
>>>>>>
>>>>>> Thanks,
>>>>>> ferruh
>>>>>
>>>>
>>