[RFC PATCH v4 2/4] dpll: Add DPLL framework base functions

[Jiri Pirko]

Fri, Dec 02, 2022 at 03:54:33PM CET, arkadiusz.kubalewski at intel.com wrote:
>>From: Jiri Pirko <jiri at resnulli.us>
>>Sent: Friday, December 2, 2022 1:40 PM
>>
>>Fri, Dec 02, 2022 at 12:27:35PM CET, arkadiusz.kubalewski at intel.com wrote:
>>>>From: Jiri Pirko <jiri at resnulli.us>
>>>>Sent: Wednesday, November 30, 2022 5:37 PM Tue, Nov 29, 2022 at
>>>>10:37:22PM CET, vfedorenko at novek.ru wrote:
>>>>>From: Vadim Fedorenko <vadfed at fb.com>
>>
>>[...]
>>
>>
>>>>>+static int
>>>>>+dpll_msg_add_pin_netifindex(struct sk_buff *msg, const struct
>>>>dpll_pin_attr *attr)
>>>>>+{
>>>>>+	unsigned int netifindex; // TODO: Should be u32?
>>>>>+
>>>>>+	if (dpll_pin_attr_netifindex_get(attr, &netifindex))
>>>>>+		return 0;
>>>>>+	if (nla_put_u32(msg, DPLLA_PIN_NETIFINDEX, netifindex))
>>>>
>>>>I was thinking about this. It is problematic. DPLL has no notion of
>>>>network namespaces. So if the driver passes ifindex, dpll/user has no
>>>>clue in which network namespace it is (ifindexes ovelay in multiple
>>>>namespaces).
>>>>
>>>>There is no easy/nice solution. For now, I would go without this and
>>>>only have linkage the opposite direction, from netdev to dpll.
>>>
>>>Well, makes sense to me.
>>>Although as I have checked `ip a` showed the same ifindex either if
>>>port was in the namespace or not.
>>
>>That is not the problem. The problem is, that you can have following two
>>netdevs with the same ifindex each in different netns.
>>1) netdev x: ifindex 8, netns ns1
>>2) netdev y: ifindex 8, netns ns2
>>
>
>OK, I now see your point what is the confusion.
>Thanks for explanation.
>But I am still not sure how to make it this way in Linux, if interface added to
>netns uses original netdev ifindex, and driver after reload receives new
>(previously unused ifindex) what would be the steps/commands to make it as you
>described? 

As I said, I don't see a way to have the ifindex exposed throught dpll
at all. I believe we should do it only the other way around. Assign
dpll_pin pointer to struct net_device and expose this over new attr
IFLA_DPLL_PIN over RT netlink.


>
>>>Isn't it better to let the user know ifindex, even if he has to iterate
>>>all the namespaces he has created?
>>
>>Definitelly not. As I showed above, one ifindex may refer to multiple
>>netdevice instances.
>>
>>
>>[...]
>>
>>
>>>>>+	DPLLA_NETIFINDEX,
>>>>
>>>>Duplicate, you have it under pin.
>>>
>>>The pin can have netifindex as pin signal source may originate there by
>>>Clock recovery mechanics.
>>>The dpll can have ifindex as it "owns" the dpll.
>>
>>DPLL is not owned by any netdevice. That does not make any sense.
>>Netdevice may be "child" of the same PCI device as the dpll instance.
>>But that's it.
>
>Sure, I won't insist on having it there, as I said, thought Maciej have seen
>a benefit with such traceability, unfortunately I cannot recall what was it.
>
>
>Thanks,
>Arkadiusz
> 
>>
>>
>>>Shall user know about it? probably nothing usefull for him, although
>>>didn't Maciej Machnikowski asked to have such traceability?