[openib-general] [CM] what happen if the path in the REQ packet (primary or alternate) is not reversible?

[Michael S. Tsirkin]

> Quoting r. Sean Hefty <mshefty at ichips.intel.com>:
> > The reason is that today there aren't many (any?) heterogeneous clusters
> > with SDR and DDR nodes, but we need to be prepared ..
> 
> Would that affect whether a path is reversible?
> 
> > I think that if the CM doesn't support non reversible paths, at least it
> > should fail to send the REQ in order to let the user understand that there
> > is a problem ...
> 
> This would be easy enough to add.  The rdma cm should also only query for 
> reversible paths.

And in particular, it seems that we need to fix SRP to set IB_SA_PATH_REC_REVERSIBLE.
Roland, does the following untested patch look right to you?

SRP must set IB_SA_PATH_REC_REVERSIBLE since that's the only kind of path
CM currently supports.

Signed-off-by: Michael S. Tsirkin <mst at mellanox.co.il>

---

diff --git a/drivers/infiniband/ulp/srp/ib_srp.c b/drivers/infiniband/ulp/srp/ib_srp.c
index 4b09147..85111b4 100644
--- a/drivers/infiniband/ulp/srp/ib_srp.c
+++ b/drivers/infiniband/ulp/srp/ib_srp.c
@@ -276,7 +276,8 @@ static int srp_lookup_path(struct srp_ta
 						   IB_SA_PATH_REC_DGID		|
 						   IB_SA_PATH_REC_SGID		|
 						   IB_SA_PATH_REC_NUMB_PATH	|
-						   IB_SA_PATH_REC_PKEY,
+						   IB_SA_PATH_REC_PKEY          |
+						   IB_SA_PATH_REC_REVERSIBLE,
 						   SRP_PATH_REC_TIMEOUT_MS,
 						   GFP_KERNEL,
 						   srp_path_rec_completion,

-- 
MST