wireless management... kernel or user space?

wireless management... kernel or user space?

[James Ketrenos]

With the various discussions going on about a generic 802.11 frame/management 
stack, and the recent thread on wireless extensions, I wanted to find out what 
people's thoughts are on where wireless management should occur -- kernel or 
user space...

Within the hostap project there exists the ability to have some AP management 
occur internal to the driver or (via the PRISM2_NO_KERNEL_IEEE80211_MGMT define) 
in a user space daemon.

Does anyone have any thoughts on how much logic should be placed into the 
generic 802.11 frame/management/control handling stack vs. pushing to a user 
space component?

If (for example) AP associate request / response policy is pushed to a user 
space component, by what mechanism does it make sense for the driver/application 
to send/receive the 802.11 frames?  Are people happy with the way Host AP 
currently does it, or is there a better way?

Other types of policy that could occur in user space would be for scan requests, 
probe response collection, AP selection to associate, etc.  What is the general 
direction the wireless networking stack should take?  Minimize kernel side 
wherever possible?

With ipw2100 (and I believe other cards) the hw/fw handles a lot of the above 
internally, so the mechanism by which the user space application communicates 
with the driver would need allow for hw/fw capabilities to be exposed so the 
user doesn't get a broken experience.

Thoughts?

James
(of ipw2100.sf.net)

Re: wireless management... kernel or user space?

[Vladimir Kondratiev]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On Saturday 19 June 2004 22:04, Jouni Malinen wrote:
> > Does anyone have any thoughts on how much logic should be placed into the
> > generic 802.11 frame/management/control handling stack vs. pushing to a
> > user space component?
>
> The ideal location of this logic implementation seems to depend a bit on
> what one is doing..
[skip]
I would add 2 points:

- - we should look carefully for timing requirements. For example, for 
association req/resp, there is no timing requirement in the standard at all; 
for auth sequence with WEP, there is maximum timeout between 2-nd and 3-rd 
frames, but it is not very strong. My point is, if there is flow that require 
timing under 10-20 ms (or even 50), it should be done in kernel.

- - there are lots of meta-information associated with self and wireless peers, 
related to rate scaling, QoS parameters, authentication details etc. I would 
say there is such entity, "peers data base". It may reside in either user 
space, or within the kernel. Logically, management entity and peers data base 
should be on the same side of kernel/user boundary. Rate scaling, if done on 
the host, is very time critical and should be done in kernel. This pushes me 
to the point that it should be in kernel.

BTW, .1X flow is pure data. It is not management one, in terms of 802.11.

Vladimir.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFA1Swpqxdj7mhC6o0RAhH3AKCL5CBpytVRs65jl8HHCI2eSbasNACgo+vZ
ta5MHMuyGgvqOtlt6LjiD+s=
=ZZIH
-----END PGP SIGNATURE-----