On 1/31/2017 3:30 PM, Pascal Mazon wrote:
> On 01/31/2017 04:28 PM, Ferruh Yigit wrote:
>> On 1/31/2017 2:23 PM, Pascal Mazon wrote:
>>> On 01/31/2017 02:06 PM, Ferruh Yigit wrote:> On 1/31/2017 9:42 AM,
>>> Pascal Mazon wrote:
>>>>> dev->data->name contains "net_tap", the device driver name.
>>>>
>>>> I see what patch does, just as a note to commit log:
>>>>
>>>> AFAIK, "dev->data->name" is device name, and for this case it is
>>>> "net_tap#", like "net_tap0", "net_tap1" ...
>>>>
>>>> "dev->data_drv_name" is the driver name which is "net_tap"
>>>
>>> Indeed, dev->data->name is the device name, looking like "net_tap#",
>>> with number increasing for each vdev.
>>> I'll put the following commit log line if that's ok:
>>>
>>> dev->data->name contains the device name, e.g. "net_tap0".
>>>
>>>>
>>>>> dev->data->dev_private->name contains the actual iface name,
>>>>> e.g. "dtap0".
>>>>
>>>> Right, I agree this is correct comparing "dev->data->name"
>>>>
>>>> But the problem is pmd->name is per eth_dev.
>>>>
>>>> If I read code correct, for multiple queue support, each queue pair will
>>>> create a tap device, so each needs a different name.
>>>>
>>>> So can't just use pmd->name. Need to create a name per queue pair, it
>>>> can be combination of pmd->name + "_" + queue_id? Or can keep a name per
>>>> queue pair, instead of eth_dev.
>>>>
>>>> What do you think?
>>>
>>> Actually that's not exactly how it goes.
>>> Adding a queue to a netdevice requires to open("/dev/net/tun") and setting
>>> TUNSETIFF (through ioctl) on the resulting fd.
>>> That's the important part: queues for a given tap device must set TUNSETIFF
>>> with a common ifreq.ifr_name (in our case, pmd->name).
>>>
>>> This is best explained in the kernel doc, there:
>>
>> Thank you for the clarification.
>> If so, why PMD keeps a fd per queue?
>>
>> Overall, patch looks good except mentioned detail in commit log.
>>
>> I suggest waiting Keith's patch, and rebase this set on top of it.
>>
>> Thanks,
>> ferruh
>>
>>>
>>> [1]
>>> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/tree/Documentation/networking/tuntap.txt#n108
>>>
>> <...>
>>
>
> I would say it is because dev->{r/t}x_pkt_burst() is done per queue.
> In these functions, we need to make our read() and write() accesses on the
> right fd considering the queue we're required to process.
If fd is same for all queues, it is possible to keep one instance in pmd
private_data, and access it from queues. I think it is confusing to have
multiple copy of it, but I see your point.
>
> I'll wait for Keith's patch, then.
Thanks again.
>
> Best regards,
>
