[RFC5040 question] RDMA/siw: R is set without an RDMA header for a Send-with-Invalidate error

Hello,

RFC 5040 Section 4.8 says that a Terminated RDMA Header must be
present when the R bit is set. Section 7.1 case 2 says that the
Terminated RDMA Header is zero for errors on incoming Send messages.

On Linux master commit
075b74841bd0065a3bda3440873c747938e69b68, an invalidation failure
returning -EACCES is classified as an RDMAP Remote Protection Error
with in_tx=0:

https://github.com/torvalds/linux/blob/075b74841bd0065a3bda3440873c747938e69b68/drivers/infiniband/sw/siw/siw_qp_rx.c#L1248-L1258

siw_send_terminate() then sets M, D, and R unconditionally:

https://github.com/torvalds/linux/blob/075b74841bd0065a3bda3440873c747938e69b68/drivers/infiniband/sw/siw/siw_qp.c#L427-L435

For the non-RREQ path, however, it appends only sizeof(struct
iwarp_send):

https://github.com/torvalds/linux/blob/075b74841bd0065a3bda3440873c747938e69b68/drivers/infiniband/sw/siw/siw_qp.c#L470-L483

That object supplies the DDP Segment Length and terminated DDP header,
but no 224-bit RDMA Read Request header. The resulting Terminate
therefore appears to advertise R=1 without carrying the corresponding
RDMA header.

Figure 10 also says that a Remote Protection Error includes an RDMA
header while listing Send with Invalidate as a possible cause, which
seems difficult to reconcile with the Section 4.8 prose and Section
7.1 case 2. Should R be cleared for the non-RREQ branch, or is another
header representation intended here?

I have not reproduced this on a live connection; this is based on
static source and wire-layout analysis.

This finding came from an academic source-level RFC consistency review.

Best regards