Re: [RESEND PATCH net-next 1/4] net: hsr: generate supervision frame without HSR tag

On Mon, Feb 01, 2021 at 01:43:43PM -0600, George McCollister wrote:

quoted

quoted

diff --git a/net/hsr/hsr_device.c b/net/hsr/hsr_device.c
index ab953a1a0d6c..161b8da6a21d 100644
--- a/net/hsr/hsr_device.c
+++ b/net/hsr/hsr_device.c

@@ -242,8 +242,7 @@ static struct sk_buff *hsr_init_skb(struct hsr_port *master, u16 proto)
       * being, for PRP it is a trailer and for HSR it is a
       * header
       */
-     skb = dev_alloc_skb(sizeof(struct hsr_tag) +
-                         sizeof(struct hsr_sup_tag) +
+     skb = dev_alloc_skb(sizeof(struct hsr_sup_tag) +
                          sizeof(struct hsr_sup_payload) + hlen + tlen);

Question 1: why are you no longer allocating struct hsr_tag (or struct prp_rct,
which has the same size)?

Because the tag is no longer being included in the supervisory frame
here. If I understand correctly hsr_create_tagged_frame and
prp_create_tagged_frame will create a new skb with HSR_HLEN added
later.

I'm mostly doing static analysis of the code, which makes everything
more difficult and also my review more inaccurate. I'll try to give your
patches more testing when reviewing further, but I just got stuck into
trying to understand them first.

So your change makes fill_frame_info classify supervision frames as
skb_std instead of skb_hsr or skb_prp. The tag is added only in
hsr_create_tagged_frame right before dispatch to the egress port.

But that means that there are places like for example
hsr_handle_sup_frame which clearly don't like that: it checks whether
there's a tagged skb in either frame->skb_hsr or frame->skb_prp, but not
in frame->skb_std, so it now does basically nothing.

Don't we need hsr_handle_sup_frame?

quoted

In hsr->proto_ops->fill_frame_info in the call path above, the skb is
still put either into frame->skb_hsr or into frame->skb_prp, but not
into frame->skb_std, even if it does not contain a struct hsr_tag.

Are you sure? My patch changes hsr_fill_frame_info and
prp_fill_frame_info not to do that if port->type is HSR_PT_MASTER
which I'm pretty certain it always is when sending supervisory frames
like this. If I've overlooked something let me know.

You're right, I had figured it out myself in the comment below where I
called it a kludge.

quoted

Also, which code exactly will insert the hsr_tag later? I assume
hsr_fill_tag via hsr->proto_ops->create_tagged_frame?

Correct.

I think it's too late, see above.

quoted

quoted

-     if (!hsr->prot_version)
-             proto = ETH_P_PRP;
-     else
-             proto = ETH_P_HSR;
-
-     skb = hsr_init_skb(master, proto);
+     skb = hsr_init_skb(master, ETH_P_PRP);

Question 2: why is this correct, setting skb->protocol to ETH_P_PRP
(HSR v0) regardless of prot_version? Also, why is the change necessary?

This part is not intuitive and I don't have a copy of the documents
where v0 was defined. It's unfortunate this code even supports v0
because AFAIK no one else uses it; but it's in here so we have to keep
supporting it I guess.
In v1 the tag has an eth type of 0x892f and the encapsulated
supervisory frame has a type of 0x88fb. In v0 0x88fb is used for the
eth type and there is no encapsulation type. So... this is correct
however I compared supervisory frame generation before and after this
patch for v0 and I found a problem. My changes make it add 0x88fb
again later for v0 which it's not supposed to do. I'll have to fix
that part somehow.

Step 1: Sign up for HSR maintainership, it's currently orphan
Step 2: Delete HSRv0 support
Step 3: See if anyone shouts, probably not
Step 4: Profit.

quoted

Why is it such a big deal if supervision frames have HSR/PRP tag or not?

Because if the switch does automatic HSR/PRP tag insertion it will end
up in there twice. You simply can't send anything with an HSR/PRP tag
if this is offloaded.

When exactly will your hardware push a second HSR tag when the incoming
packet already contains one? Obviously for tagged packets coming from
the ring it should not do that. It must be treating the CPU port special
somehow, but I don't understand how.