3.2.96-rc1 review patch. If anyone has any objections, please let me know.
------------------
From: Mike Marciniszyn mike.marciniszyn@intel.com
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 --- --- a/drivers/infiniband/hw/qib/qib_rc.c +++ b/drivers/infiniband/hw/qib/qib_rc.c @@ -365,7 +365,7 @@ int qib_make_rc_req(struct qib_qp *qp) case IB_WR_RDMA_WRITE: if (newreq && !(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) qp->s_lsn++; - /* FALLTHROUGH */ + goto no_flow_control; case IB_WR_RDMA_WRITE_WITH_IMM: /* If no credit, return. */ if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT) && @@ -373,6 +373,7 @@ int qib_make_rc_req(struct qib_qp *qp) qp->s_flags |= QIB_S_WAIT_SSN_CREDIT; goto bail; } +no_flow_control: ohdr->u.rc.reth.vaddr = cpu_to_be64(wqe->wr.wr.rdma.remote_addr); ohdr->u.rc.reth.rkey =