CM REQs cannot be successfully retried, because a new pv_cm_id is
created for each request, without checking if one already exists.
By checking if an id exists before creating one, the bug is fixed.
This bug can be provoked by running an RDMA CM user-land application,
but inserting a five seconds delay before the rdma_accept() call on
the passive side. This delay is larger than the default CMA timeout,
and triggers a retry from the active side. The retried REQ will use
another pv_cm_id (the cm_id on the wire). This confuses the CM
protocol and two REJs are sent from the passive side.
Here is an excerpt from ibdump running without the patch:
3.285092 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello)
7.382711 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello)
7.382861 LID: 4 -> LID: 4 InfiniBand 290 CM: ConnectReject
7.387644 LID: 4 -> LID: 4 InfiniBand 290 CM: ConnectReject
and here is the same with bug fix applied:
3.251010 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello)
7.349387 LID: 4 -> LID: 4 SDP 290 CM: ConnectRequest(SDP Hello)
8.258443 LID: 4 -> LID: 4 SDP 290 CM: ConnectReply(SDP Hello)
8.259890 LID: 4 -> LID: 4 InfiniBand 290 CM: ReadyToUse
Suggested-by: Venkat Venkatsubra <venkat.x.venkatsubra@oracle.com>
Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com>
Reported-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Tested-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Acked-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Doug Ledford <dledford@redhat.com>
4542e3c79a