Re: [net-next v41] mctp pcc: Implement MCTP over PCC Transport

[Jeremy Kerr]

Hi Adam,

> > Adding the skb_linearize() there is unnecessary, and creates ambiguity
> > about the structure of the skbs that we're dealing with in that path.
> > 
> OK, I get it now.  I was thinking that the Fragmentation was a likely
> part of the progression, as default MCTP packets are so small.

OK, beware of potential confusion; there are two separate concepts here,
"MCTP fragmentation" and "skbs that have fragments".

"MCTP fragmentation" is implemented by the MCTP core, and splits large
MCTP messages into smaller packets; each packet being a separate skb,
with its own MCTP header.

Your transport driver does not need to care about this; it is only
concerned with transferring those individual skbs.

"skbs that have fragments" refers to the possibility that the data
represented by a skb may be held in scatter-gather format, rather than
(or in addition to) the skb->data buffer.

Your transport driver does not need to care about this either, as it is
not claiming to support fragmented skbs, so all of the skbs that reach
->ndo_start_xmit() will be linear.

> Yeah, I saw those, and was processing through them.  I have found 
> similar type issues running against Codex.
> 
> Some of the issues I found are due to the interactions with the PCC 
> layer, and need to be addressed there.  It leads to some edge conditions 
> that are, for me, impossible to produce right now. There is one change 
> it suggests which I thought I had already made.
> 
> If those changes are going to cause significatnt changes, I will 
> probably also integrate your suggestion on how to deal with IRQ-safe 
> stats posting.

OK, sounds good.

Cheers,


Jeremy