RDMA/cxgb3: Turn off RX coalescing for iWARP connections

The HW by default has RX coalescing on.  For iWARP connections, this
causes a 100ms delay in connection establishement due to the ingress
MPA Start message being stalled in HW.  So explicitly turn RX
coalescing off when setting up iWARP connections.

This was causing very bad performance for NP64 gather operations using
Open MPI, due to the way it sets up connections on larger jobs.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Cc: <stable@kernel.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

diff --git a/drivers/infiniband/hw/cxgb3/iwch_cm.c b/drivers/infiniband/hw/cxgb3/iwch_cm.c
index d88077a..13c8887 100644 (file)
--- a/drivers/infiniband/hw/cxgb3/iwch_cm.c
+++ b/drivers/infiniband/hw/cxgb3/iwch_cm.c
@@ -463,7 +463,8 @@ static int send_connect(struct iwch_ep *ep)
            V_MSS_IDX(mtu_idx) |
            V_L2T_IDX(ep->l2t->idx) | V_TX_CHANNEL(ep->l2t->smt_idx);
        opt0l = V_TOS((ep->tos >> 2) & M_TOS) | V_RCV_BUFSIZ(rcv_win>>10);
-       opt2 = V_FLAVORS_VALID(1) | V_CONG_CONTROL_FLAVOR(cong_flavor);
+       opt2 = F_RX_COALESCE_VALID | V_RX_COALESCE(0) | V_FLAVORS_VALID(1) |
+              V_CONG_CONTROL_FLAVOR(cong_flavor);
        skb->priority = CPL_PRIORITY_SETUP;
        set_arp_failure_handler(skb, act_open_req_arp_failure);
 
@@ -1280,7 +1281,8 @@ static void accept_cr(struct iwch_ep *ep, __be32 peer_ip, struct sk_buff *skb)
            V_MSS_IDX(mtu_idx) |
            V_L2T_IDX(ep->l2t->idx) | V_TX_CHANNEL(ep->l2t->smt_idx);
        opt0l = V_TOS((ep->tos >> 2) & M_TOS) | V_RCV_BUFSIZ(rcv_win>>10);
-       opt2 = V_FLAVORS_VALID(1) | V_CONG_CONTROL_FLAVOR(cong_flavor);
+       opt2 = F_RX_COALESCE_VALID | V_RX_COALESCE(0) | V_FLAVORS_VALID(1) |
+              V_CONG_CONTROL_FLAVOR(cong_flavor);
 
        rpl = cplhdr(skb);
        rpl->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_FORWARD));