On 2018-06-15 17:49, Johannes Berg wrote:
On Mon, 2018-05-28 at 19:04 +0530, Sriram R wrote:
Per-rate/per-station statistics can be desirable to have but they're
quite expensive (keeping the four counters for each rate would take
close to 4k of memory per station only for VHT MCSes for a moderately
capable VHT chip (with 2 spatial streams and 80MHz support) so it's
not a good idea to keep all of this in the kernel.

Instead, this API provides a way for interested clients in userspace
to subscribe to such statistics. When supported by a driver, it can
then start collecting the data only when subscribers exist. To avoid
the kernel's data collection becoming too big, it can send out the
data at any point in time, for example to limit the counters to u16
internally and send it out when they're close to reaching the limit,
or to keep a hash table and sending it out when too many collisions
occur. Userspace can then keep track of the full state.

Based on below implementation by Johannes Berg <johan...@sipsolutions.net>
http://thread.gmane.org/gmane.linux.kernel.wireless.general/133172
with following changes,

Err, that's pretty much an exact copy of my patch ...

You should keep the author attribution and signed-off-by, and if you
want your work to be separately attributed then just put the changes you
made into (a) separate patch(es).
Sure! I'll correct this in my next revision.

 1. Allow rx bytes stats to be collected

This seems OK.

 2. Rate info sent to userspace is encoded into 32bit value

This I don't like at all. It's fine for the internal representation, and I know I did something like that in mac80211 (though I guess at the time
it was only 16 bit), but it's not OK for the userspace representation.

I just merged patches to support HE in the userspace representation, and
that's properly taken into account in struct rate_info which I had here
in my version of the patch.
Okay Johannes. I'll revert this in my next revision.

johannes

Reply via email to