RDMA/qedr: Fix wmb usage in qedr

This patch comes as a result of Sinan Kaya's work and the decision that
writel() must be a strong enough barrier for DMA.

wmb usages in qedr driver have either been removed where they were there
only to order DMA accesses, and replaced with smp_wmb and comments for the
places that the barrier was there for SMP reasons.

Fixes: 561e5d48968b ("RDMA/qedr: eliminate duplicate barriers on weakly-ordered archs")
Signed-off-by: Michal Kalderon <Michal.Kalderon@cavium.com>
Signed-off-by: Ariel Elior <Ariel.Elior@cavium.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c
index 1835dc9eb3e3992deb337e8f54ccf4ba12600d9c..26a698927d2e4a00ffe51602f0fb1ab9dc446103 100644 (file)
--- a/drivers/infiniband/hw/qedr/verbs.c
+++ b/drivers/infiniband/hw/qedr/verbs.c
@@ -804,8 +804,6 @@ static inline void qedr_init_cq_params(struct qedr_cq *cq,
 
 static void doorbell_cq(struct qedr_cq *cq, u32 cons, u8 flags)
 {
-       /* Flush data before signalling doorbell */
-       wmb();
        cq->db.data.agg_flags = flags;
        cq->db.data.value = cpu_to_le32(cons);
        writeq(cq->db.raw, cq->db_addr);
@@ -1812,8 +1810,7 @@ static int qedr_update_qp_state(struct qedr_dev *dev,
                         */
 
                        if (rdma_protocol_roce(&dev->ibdev, 1)) {
-                               wmb();
-                               writel_relaxed(qp->rq.db_data.raw, qp->rq.db);
+                               writel(qp->rq.db_data.raw, qp->rq.db);
                                /* Make sure write takes effect */
                                mmiowb();
                        }
@@ -3198,9 +3195,16 @@ int qedr_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
         * vane. However this is not harmful (as long as the producer value is
         * unchanged). For performance reasons we avoid checking for this
         * redundant doorbell.
+        *
+        * qp->wqe_wr_id is accessed during qedr_poll_cq, as
+        * soon as we give the doorbell, we could get a completion
+        * for this wr, therefore we need to make sure that the
+        * memory is updated before giving the doorbell.
+        * During qedr_poll_cq, rmb is called before accessing the
+        * cqe. This covers for the smp_rmb as well.
         */
-       wmb();
-       writel_relaxed(qp->sq.db_data.raw, qp->sq.db);
+       smp_wmb();
+       writel(qp->sq.db_data.raw, qp->sq.db);
 
        /* Make sure write sticks */
        mmiowb();
@@ -3286,8 +3290,14 @@ int qedr_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 
                qedr_inc_sw_prod(&qp->rq);
 
-               /* Flush all the writes before signalling doorbell */
-               wmb();
+               /* qp->rqe_wr_id is accessed during qedr_poll_cq, as
+                * soon as we give the doorbell, we could get a completion
+                * for this wr, therefore we need to make sure that the
+                * memory is update before giving the doorbell.
+                * During qedr_poll_cq, rmb is called before accessing the
+                * cqe. This covers for the smp_rmb as well.
+                */
+               smp_wmb();
 
                qp->rq.db_data.data.value++;