Re: [PATCH V3 5/6] vDPA: answer num of queue pairs = 1 to userspace when VIRTIO_NET_F_MQ == 0

[Si-Wei Liu <si-wei.liu@oracle.com>]

On 7/27/2022 2:01 AM, Michael S. Tsirkin wrote:
> On Wed, Jul 27, 2022 at 12:50:33AM -0700, Si-Wei Liu wrote:
>>
>> On 7/26/2022 11:01 PM, Michael S. Tsirkin wrote:
>>> On Wed, Jul 27, 2022 at 03:47:35AM +0000, Parav Pandit wrote:
>>>>> From: Zhu, Lingshan <lingshan.zhu@intel.com>
>>>>> Sent: Tuesday, July 26, 2022 10:53 PM
>>>>>
>>>>> On 7/27/2022 10:17 AM, Parav Pandit wrote:
>>>>>>> From: Zhu, Lingshan <lingshan.zhu@intel.com>
>>>>>>> Sent: Tuesday, July 26, 2022 10:15 PM
>>>>>>>
>>>>>>> On 7/26/2022 11:56 PM, Parav Pandit wrote:
>>>>>>>>> From: Zhu, Lingshan <lingshan.zhu@intel.com>
>>>>>>>>> Sent: Tuesday, July 12, 2022 11:46 PM
>>>>>>>>>> When the user space which invokes netlink commands, detects that
>>>>>>> _MQ
>>>>>>>>> is not supported, hence it takes max_queue_pair = 1 by itself.
>>>>>>>>> I think the kernel module have all necessary information and it is
>>>>>>>>> the only one which have precise information of a device, so it
>>>>>>>>> should answer precisely than let the user space guess. The kernel
>>>>>>>>> module should be reliable than stay silent, leave the question to
>>>>>>>>> the user space
>>>>>>> tool.
>>>>>>>> Kernel is reliable. It doesn’t expose a config space field if the
>>>>>>>> field doesn’t
>>>>>>> exist regardless of field should have default or no default.
>>>>>>> so when you know it is one queue pair, you should answer one, not try
>>>>>>> to guess.
>>>>>>>> User space should not guess either. User space gets to see if _MQ
>>>>>>> present/not present. If _MQ present than get reliable data from kernel.
>>>>>>>> If _MQ not present, it means this device has one VQ pair.
>>>>>>> it is still a guess, right? And all user space tools implemented this
>>>>>>> feature need to guess
>>>>>> No. it is not a guess.
>>>>>> It is explicitly checking the _MQ feature and deriving the value.
>>>>>> The code you proposed will be present in the user space.
>>>>>> It will be uniform for _MQ and 10 other features that are present now and
>>>>> in the future.
>>>>> MQ and other features like RSS are different. If there is no _RSS_XX, there
>>>>> are no attributes like max_rss_key_size, and there is not a default value.
>>>>> But for MQ, we know it has to be 1 wihtout _MQ.
>>>> "we" = user space.
>>>> To keep the consistency among all the config space fields.
>>> Actually I looked and the code some more and I'm puzzled:
>>>
>>>
>>> 	struct virtio_net_config config = {};
>>> 	u64 features;
>>> 	u16 val_u16;
>>>
>>> 	vdpa_get_config_unlocked(vdev, 0, &config, sizeof(config));
>>>
>>> 	if (nla_put(msg, VDPA_ATTR_DEV_NET_CFG_MACADDR, sizeof(config.mac),
>>> 		    config.mac))
>>> 		return -EMSGSIZE;
>>>
>>>
>>> Mac returned even without VIRTIO_NET_F_MAC
>>>
>>>
>>> 	val_u16 = le16_to_cpu(config.status);
>>> 	if (nla_put_u16(msg, VDPA_ATTR_DEV_NET_STATUS, val_u16))
>>> 		return -EMSGSIZE;
>>>
>>>
>>> status returned even without VIRTIO_NET_F_STATUS
>>>
>>> 	val_u16 = le16_to_cpu(config.mtu);
>>> 	if (nla_put_u16(msg, VDPA_ATTR_DEV_NET_CFG_MTU, val_u16))
>>> 		return -EMSGSIZE;
>>>
>>>
>>> MTU returned even without VIRTIO_NET_F_MTU
>>>
>>>
>>> What's going on here?
>>>
>>>
>> I guess this is spec thing (historical debt), I vaguely recall these fields
>> are always present in config space regardless the existence of corresponding
>> feature bit.
>>
>> -Siwei
> Nope:
>
> 2.5.1  Driver Requirements: Device Configuration Space
>
> ...
>
> For optional configuration space fields, the driver MUST check that the corresponding feature is offered
> before accessing that part of the configuration space.
Well, this is driver side of requirement. As this interface is for host 
admin tool to query or configure vdpa device, we don't have to wait 
until feature negotiation is done on guest driver to extract vdpa 
attributes/parameters, say if we want to replicate another vdpa device 
with the same config on migration destination. I think what may need to 
be fix is to move off from using .vdpa_get_config_unlocked() which 
depends on feature negotiation. And/or expose config space register 
values through another set of attributes.

-Siwei