Re: [RFC] Introduce NL80211_IFTYPE_MLO_LINK_DEVICE for MLO link (IEEE 802.11be)

[usdutt@codeaurora.org]

On 2020-12-14 14:55, Arend van Spriel wrote:
> On 09-12-2020 16:36, Johannes Berg wrote:
>> Hi,
>> 
>>> Multi-link support is introduced in 802.11be specification.
>> [...]
>> 
>> I'll definitely have to take a closer look at this and the spec and
>> think about it - but a couple of quick comments below.

Thanks for your comments. Please allow me to address your comments in 
the subsequent patches.
Our main intention through this RFC is to get the view on the proposal 
to introduce a new NL80211_IFTYPE for an MLO Link.
Please suggest on this.


> Did not get my hands on the spec yet, but just firing random thoughts.
> This multi-link support sounds a bit like bonding. Could that be
> leveraged for wireless? I wonder why there is a need for a non-netdev
> interface. Does networking layer need to be aware of the multi-link
> for scheduling traffic to it? Is there one driver per multi-link or
> per radio?

The following is what the Bonding driver / interface does for the 
multiple network interfaces.
1. Bonding driver aggregates multiple network interfaces into a single 
logical bonded interface.
2. Each network interface is called as a slave in the bonding driver.
3. It’s the Bonding driver that schedules the traffic across these 
slaves.
4. Bonding interface operates on one of the following modes ( mentioning 
for a quick reference ).
        balance-rr : Round-robin policy: Transmit packets in sequential 
order from the first available slave through the last.
        active-backup : Only one slave in the bond is active. A different 
slave becomes active if, and only if, the active slave fails.
        balance-xor : Transmit based on the selected transmit hash 
policy.
        balance-tlb : channel bonding that does not require any special 
switch support. The outgoing traffic is distributed according to the 
current load
        balance-alb : includes balance-tlb plus receive load balancing.
5. With the above modes, the scheduling of the traffic is determined 
with in the kernel / bonding driver and have the host driver / firmware 
follow it.

The following are the factors for not considering the MLO Link as a 
bonding interface and further to propose this RFC
(new NL80211_IF_TYPE -> NL80211_IFTYPE_MLO_LINK_DEVICE).

1. For MLO, our intention is to allow the traffic scheduling (among the 
MLO links) to be closer to the lower layers in
         the driver/firmware, etc, unlike the bonding driver where kernel 
schedules the traffic.
2. If MLO Link has to use a bonding interface, each MLO link has to be 
associated with a network interface and thus each
         of the MLO link has to be of NL80211_IF_TYPE  -> 
NL80211_IFTYPE_STATION.
3. This further means that NL80211_CONNECT / NL80211_ASSOCIATE needs to 
get triggered on each of this MLO Link of the
         MLO connection.
4. Our intention is to consider the MLO station as a single network 
interface . More specifically, have a single
         NL80211_CONNECT / NL80211_CMD_ASSOCIATE triggered for each MLO 
connection ( MLD + MLO Links).
5. The reason we had to introduce a new NL80211_IF_TYPE -> 
NL80211_IFTYPE_MLO_LINK_DEVICE is to have the corresponding
        link connection specific information ( e.g., Connection State, 
Connected BSS Info, Connection Frequency/properties, etc),
        for each wdev / MLO Link and at the same  time have a single 
connect / association request to the driver / AP.
6. Further, these NL80211_IFTYPE_MLO_LINK_DEVICE link devices shall 
represent an MLO Link and attribute to a network interface ( MLD ) of an
       MLO connection.
7. The current proposal to introduce a new NL80211_IF_TYPE should 
support the architectures of MLO links across the multiple wiphy
         interfaces or on the same wiphy interface. These wiphy 
interfaces can be from different drivers too.

Please review the above.

> 
> [...]
> 
>>>   	struct sk_buff *msg;
>>>   	void *hdr;
>>> +	struct nlattr *nested, *nested_mlo_links;
>>> +	struct cfg80211_mlo_link_device_params *mlo_link;
>>> +	int i = 0;
>>>     	msg = nlmsg_new(100 + cr->req_ie_len + cr->resp_ie_len +
>>>   			cr->fils.kek_len + cr->fils.pmk_len +
>>> -			(cr->fils.pmkid ? WLAN_PMKID_LEN : 0), gfp);
>>> +			(cr->fils.pmkid ? WLAN_PMKID_LEN : 0) +
>>> +			(cr->n_mlo_links ? cr->n_mlo_links * 32 : 0), gfp);
>> 
>> 32 probably should be some NLA_SIZE or something constant?
> 
> Probably sizeof(*mlo_link) ?
> 
>> There's also no point in the ternary operator since 0 * 32 is 0 :)
>> 
>>>   /* Consumes bss object one way or another */
>>> @@ -833,7 +849,9 @@ void cfg80211_connect_done(struct net_device 
>>> *dev,
>>>   	ev = kzalloc(sizeof(*ev) + (params->bssid ? ETH_ALEN : 0) +
>>>   		     params->req_ie_len + params->resp_ie_len +
>>>   		     params->fils.kek_len + params->fils.pmk_len +
>>> -		     (params->fils.pmkid ? WLAN_PMKID_LEN : 0), gfp);
>>> +		     (params->fils.pmkid ? WLAN_PMKID_LEN : 0) +
>>> +		     (params->n_mlo_links ? params->n_mlo_links *
>>> +		      sizeof(struct cfg80211_mlo_link_device_params) : 0), gfp);
>> 
>> same here, no need for the ternary
>> 
>> 
>> It feels strangely asymmetric to have stop and no start ... but I 
>> guess
>> that's the part where I need to think about it and look a bit at how 
>> it
>> all works :)
> 
> Had the same feeling in my gutt (wherever that is ;-)

Sure. Will take this point. In fact, this was documented in the commit 
message to consider the start (say start_mlo_link), but do not want the 
drivers to depend on this trigger for starting the MLO link.
For example, the host drivers with SME can as well start the MLO Link 
after getting the Assoc response for MLO links from the AP.


> 
> Regards,
> Arend