On 3/27/2024 8:37 PM, Johannes Berg wrote:
On Wed, 2024-03-27 at 08:02 -0700, Jeff Johnson wrote:
I'm also imagining that we change the API from cfg80211 to the drivers
to get the *link* STA information, and do the summing up and/or "best"
selection there in cfg80211 itself. However, I am prepared to accept the
possibility that we may do _both_ in the API, if not all drivers can
even do all of the statistics per link. We should probably still have
the link STAs in the statistics in nl80211, but then they may not be
populated?
First remember that there are a lot of statistics, and each driver is free to
return as many or as few as they support, indicating the ones they are
returning using the "filled" bitmap.
Yes, I'd think we want to use the same data structure for both, though
setting something in *both* links and *mld* would (should) be an error.
The statistics can be populated by driver or mac80211.(say tx retries,
tx packets etc)
So we should also change the existing stats update in mac80211 on link
STA basis instead of deflink?
I would expect MLO-capable drivers to
provide the same information on a per-link basis that they previously provided
on a per-interface basis, but the "filled" bitmap leaves that to the drivers.
Unless we don't actually ask the drivers at the MLD level if (the
station is an MLD). But I suspect we will have to do both, ask for MLD-
level stats and link-level stats.
But I think a fundamental question needs to be answered: To what extent do we
need to support legacy userspace applications that are not MLO-aware?
I have no idea who even uses this and how :-) But I guess things like
NetworkManager might, even just to show some signal strength values
etc.?
My expectation is that MLO-aware applications only need the per-link
information, and from that they can derive their own "aggregate" of the
per-link information.
Agree, though it'll be a long time until all applications are MLO-aware?
Unless there aren't many using it, but ...
So to support that we'd need per-link nesting of the
associated attributes.
Sure, that's a given.
And if we don't need to support legacy userspace we
could completely remove populating the top-level statistic attributes. Non-MLO
interfaces would have a single link nest that contains the same information
that is now in the top-level of the NL message.
But if we need to support legacy userspace then we would indeed need to
continue to populate those top-level attributes with some form of aggregate data.
I think we probably have to.
And even for the MLO-aware case there is the issue of how do we want to handle
the case that links may come and go, and hence aggregate counters would appear
to have huge fluctuations in values when links are added or removed if the
aggregate values are only calculated by adding the values from the
currently-attached links.
Good point, when they're really removed we'd want to probably keep that
value as a bias for the MLD-level stats?
ok. Then the statistics value in MLD STA would be bias + summed up value
of currently alive links?
On the same line , ethtool stats(*interface level stats*) in
mac80211(ieee80211_get_stats())
computes the stats by summing up the current STA statistics.
Here stations can come and go and the ethtool stats may not reflect the
total packets transmitted or received by the interface.
It just reflects the summed up value of current alive stations.
Since these two problems are similar (ethtool stats and MLD stats
calculation),
would like to understand what type of value would be more useful to user?
johannes
