On Tue, 2008-12-02 at 18:07 -0800, Inaky Perez-Gonzalez wrote:

On Thursday 27 November 2008, Johannes Berg wrote:

quoted

along, more functionality was moved to the host (yes, you say this
won't happen with wimax, but I think it will, eventually, if wimax
gets to be popular enough. never say never), and wext got more

Agree, never say never; but these are the reasons why I highly doubt
it'll ever happen:

 - WiMAX spectrum is heavily regulated

Well, to be fair 3.5GHz is "lightly regulated" here in the US, you just
have to register your base station with the FCC and as long as nobody
else has done so in your area, you win.  The Intel parts don't support
3.5 GHz, but I'm sure somebody elses parts will in the future.

TowerStream has already deployed 3.5GHz fixed WiMAX gear here in Boston,
though they are doing fixed access/T1-replacement for businesses, not a
large consumer-focused mobile wimax network like XOHM/Clear.

Dan

 - Operators won't allow it. They want tightly safe, certified and
   tested firmwares/devices that adhere to the standard so they can
   make sure devices don't mess up a single hertz as that means lost
   money, especially with a radio that can beam over miles.

 - Certification (required) of stacks make it almost
   impossible to maintain a soft implementation.

quoted

messy. Wext even was defining actual operations, as undefined as
they often were, you're not even doing that.

Why would I define them when I don't even know how they are going to
look like? It is not as simple as it looks.

The basics that aren't covered in current's kernel API are scan,
connect, disconnect. Disconnect is simple, but scan & connect? very
much impossible to tell what common ground is going to be until more
devices show up.

For scanning, some devices require to be told exactly where to scan in
(as in which combination of band, fft width and coloring of the
band). Some others don't. Then of course, the scan results might be
operators? Network Service Providers? Network Access Providers?  base
station IDs? how do you stitch'em together? You need information to
map from one to the other, and that is device specific depending on at
which level they work. How to stich that information together depends
on the network too (OMA-DM and provisining information help to compose
this). If it is done at the device/firmware level or at the host level
is device specific.

Connect has exactly the same levels of issues as scan: what do I
connect to? A base station? a NAP or an NSP?

So back to the original question: I have no information to define such
an interface at low level, so I am not defining it. Simple :/

quoted

I don't think you can say it's like cfg80211, cfg80211 in fact does a
lot more than the wimax "stack", it can validate parameters for
example, and provides actual operations (key operations, peer
operations, scanning [soon], ...) rather than just a transport to
the driver. As I've said on the relevant patch, I'd love to see the
wimax stack take more of that direction.

If it has to happen, it will happen eventually. If it is a stack or
not as of now and just a simple control interface, not that I really
mind. It's just a name.

quoted

I don't see that as much of a problem. Many of the commands you have
defined for the i2400m driver are very generic, for example:

As I said above, it is not so simple. They are not as generic as they
look. 

quoted

This is the biggest issue I see here. I don't see how the stack
helps anyone implement a driver for a new device. Sure, they won't
have to come up with a new transport, write less lines of generic
netlink code, but ultimately that's not the hard part, the hard part
is getting the relevant operations right etc.

I agree -- and that's the whole reason why the "stack" is
there--because even if it doesn't look so hard, it's just a
pain. 

This way someone comes in with a need for a driver and then we can
start deciding how to cut the scan & connect interfaces. Maybe it
makes no sense to do it at the kernel level and it has to be done at
user space, as the i2400m does it; then the kernel API stays the
same...until another driver comes in...etc. Don't fix it if it ain't
broke. 

quoted

This really means you're putting the actual "driver", the piece that
does the hardware abstraction, into userspace. And in a binary daemon
even, afaict. This was quickly shot down with ipw3945/4965, not sure why
nobody has cared here so far. Maybe because you're actually planning to
open source that part.

Nope. I am putting the part that knows how to scan and connect in user 
space because it does not belong in kernel space. It is big and complex, 
needs permanent storage, requires complex crypto code and can really
use a OMA-DM client to communicate with the network.

Not a binary, btw. Currently the supplicant is a binary, but that will
change. The OMA-DM client daemon is also a binary as of now and we
are still thinking how to fix that situation, as there are no open
source equivalents. Luckily, it is kind of optional.

quoted

Couldn't the stack provide more functionality here? Somewhere else
you speak of using ethernet vs. rawip, couldn't the stack do that
translation, possibly even allowing both rawip and ethernet to
coexist, or be switchable at runtime if you have a working dhcp
client?

The stack is a control plane, not a data plane. Each driver implements
data frame passing the way that is most optimal to the device it
implements support for.

There is not good reason (as of now) that justifies the extra overhead
that would be introduced by the "stack" doing it. 

The only reason I would see that happening would be a mac80211 like
case, where the stack is doing packet handling in a hypothetical
soft-mac implementation. As I said above, I highly doubt that
scenario.

Thanks,