Hi,

On Wed, Sep 21, 2022 at 11:46 AM Miquel Raynal
[off-list ref] wrote:

Hi Alexander,

aahringo@redhat.com wrote on Thu, 8 Sep 2022 20:41:14 -0400:

quoted

Hi,

On Thu, Sep 8, 2022 at 3:37 AM Miquel Raynal [off-list ref] wrote:

quoted

Hi Alexander,

aahringo@redhat.com wrote on Wed, 7 Sep 2022 21:40:13 -0400:

quoted

Hi,

On Mon, Sep 5, 2022 at 4:34 PM Miquel Raynal [off-list ref] wrote:

quoted

Hello,

A third version of this series, dropping the scan patches for now
because before we need to settle on the filtering topic and the
coordinator interface topic. Here is just the filtering part, I've
integrated Alexander's patches, as well as the atusb fix. Once this is
merge there are a few coordinator-related patches, and finally the
scan.

I think we have a communication problem here and we should talk about
what the problems are and agree on a way to solve them.

The problems are:

1. We never supported switching from an operating phy (interfaces are
up) into another filtering mode.

In the trigger scan path there is a:

        mlme_op_pre() // stop Tx
        drv_stop() // stop Rx
        synchronize_net()
        drv_start(params) // restart Rx with another hw filtering level

Okay, that's looking good.

quoted

quoted

2. Scan requires to be in "promiscuous mode" (according to the
802.15.4 spec promiscuous mode). We don't support promiscuous mode
(according to the 802.15.4 spec promiscuous mode). We "can" however
use the currently supported mode which does not filter anything
(IEEE802154_FILTERING_NONE) when we do additional filtering in
mac802154. _But_ this is only required when the phy is scanning, it
will also deliver anything to the upper layers.

This patch-series tries to do the second thing, okay that's fine. But
I thought this should only be done while the phy is in "scanning
mode"?

I don't understand what's wrong then. We ask for the "scan mode"
filtering level when starting the scan [1] and we ask for the normal
filtering level when it's done/aborted [2] [3].

There is no problem with that. There is for me a problem with the
receive path when certain filtering levels are active.

quoted

[1] https://github.com/miquelraynal/linux/blob/wpan-next/scan/net/mac802154/scan.c#L326
[2] https://github.com/miquelraynal/linux/blob/wpan-next/scan/net/mac802154/scan.c#L55

quoted

The other receive path while not in promiscuous mode
(phy->filtering == IEEE802154_FILTERING_4_FRAME_FIELDS) should never
require any additional filtering. I somehow miss this point here.

Maybe the drv_start() function should receive an sdata pointer. This way
instead of changing the PHY filtering level to what has just be asked
blindly, the code should look at the filtering level of all the
interfaces up on the PHY and apply the lowest filtering level by
hardware, knowing that on a per interface basis, the software will
compensate.

It should work just fine because local->phy->filtering shows the actual
filtering level of the PHY while sdata->requested_filtering shows the
level of filtering that was expected on each interface. If you don't
like the idea of having a mutable sdata->requested_filtering entry, I
can have an sdata->base_filtering which should only be set by
ieee802154_setup_sdata() and an sdata->expected_filtering which would
reflect what the mac expects on this interface at the present moment.

From my view is that if we disable address filters (all filtering
modes except IEEE802154_FILTERING_4_FRAME_FIELDS) we never can call
netif_receive_skb(). This patch series tries to "compensate" the
missing filtering on phy which is fine only to handle things related
for the scan operation but nothing else.

The reason why we can't call netif_receive_skb() is because we don't
have ackknowledge handling, whereas for scanning we ignore ack frames
and that's why we don't need it.

I've digested all of that right before the conference so I think I now
understand all your fears regarding the possible absence of ACK in
certain situations. I even share them now.

quoted

quoted

quoted

For 1), the driver should change the filtering mode" when we start to
"listen", this is done by the start() driver callback. They should get
all receive parameters and set up receiving to whatever mac802154,
currently there is a bit of chaos there. To move it into drv_start()
is just a workaround to begin this step that we move it at some point
to the driver. I mention 1) here because that should be part of the
picture how everything works together when the phy is switched to a
different filter level while it's operating (I mean there are running
interfaces on it which requires IEEE802154_FILTERING_4_FRAME_FIELDS)
which then activates the different receive path for the use case of
scanning (something like (phy->state & WPANPHY_SCANING) == true)?

Scanning is a dedicated filtering level per-se because it must discard
!beacon frames, that's why I request this level of filtering (which
maybe I should do on a per-interface basis instead of using the *local
poiner).

We only can do a per filter level per interface if the hardware has
support for such a thing. Currently there is one address filter and if
it's disabled we lose ackknowledge handling (as general rule), we
can't compensate by doing any additional filtering by software in this
mode.

Yes.

quoted

quoted

quoted

I am sorry, but I somehow miss the picture of how those things work
together. It is not clear for me and I miss those parts to get a whole
picture of this. For me it's not clear that those patches are going in
this direction.

Please tell me if it's more clear and if you agree with this vision. I
don't have time to draft something this week.

That's fine. We should agree on things like compensate lower filter
levels by doing additional softmac filtering to reach
IEEE802154_FILTERING_4_FRAME_FIELDS filtering from others because we
will lose AACK handling. It is only fine to do that in mac802154
receive path but don't deliver it to the upper layer.

So actually, if I try to summarize the situation.

I've tried to make several different subif working on a single PHY.
Unfortunately, today there is only one address filter per PHY, so

correct. But you are assuming hardware which is currently around. As I
said atusb is one candidate to make a more clever hardware because it
has a co-processor which eventually could handle ack handling if done
right.

disabling the address filter on one interface would also disable it on
the other, leading to the ACKs not being handled anymore, which we
cannot afford.

For hardware which has such limitation of one address filter yes.

So overall I guess we should settle on how we want to handle the
situation. I propose, to move forward, that we continue to assume that
it is *not* possible to have several different interfaces running at the
same time on a single PHY. This involves dropping all the "software
address filtering" which I proposed, but I guess that's fine.

I agree here, but don't remove code which could add such handling and
allow multiple monitors at the same time.

- Alex