IB/{qib, hfi1}: Avoid flow control testing for RDMA write operation
authorMike Marciniszyn <mike.marciniszyn@intel.com>
Tue, 22 Aug 2017 01:26:20 +0000 (18:26 -0700)
committerBen Hutchings <ben@decadent.org.uk>
Sun, 26 Nov 2017 13:51:05 +0000 (13:51 +0000)
commitff408ba4cd682a3844fd833c0281a68fec6239b3
tree1e727c7442ea919d099a8253a645f08c62dc3761
parent5d04de9a24824635e4da8da62b23a797f4774aef
IB/{qib, hfi1}: Avoid flow control testing for RDMA write operation

commit 5b0ef650bd0f820e922fcc42f1985d4621ae19cf upstream.

Section 9.7.7.2.5 of the 1.3 IBTA spec clearly says that receive
credits should never apply to RDMA write.

qib and hfi1 were doing that.  The following situation will result
in a QP hang:
- A prior SEND or RDMA_WRITE with immmediate consumed the last
  credit for a QP using RC receive buffer credits
- The prior op is acked so there are no more acks
- The peer ULP fails to post receive for some reason
- An RDMA write sees that the credits are exhausted and waits
- The peer ULP posts receive buffers
- The ULP posts a send or RDMA write that will be hung

The fix is to avoid the credit test for the RDMA write operation.

Reviewed-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
[bwh: Backported to 3.2:
 - Drop changes to hfi1
 - Adjust context]
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>
drivers/infiniband/hw/qib/qib_rc.c