xprtrdma: Reduce context switching due to Local Invalidation

Since commit ba69cd122ece ("xprtrdma: Remove support for FMR memory
registration"), FRWR is the only supported memory registration mode.

We can take advantage of the asynchronous nature of FRWR's LOCAL_INV
Work Requests to get rid of the completion wait by having the
LOCAL_INV completion handler take care of DMA unmapping MRs and
waking the upper layer RPC waiter.

This eliminates two context switches when local invalidation is
necessary. As a side benefit, we will no longer need the per-xprt
deferred completion work queue.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>

1 files changed, 102 insertions, 1 deletions

diff --git a/net/sunrpc/xprtrdma/frwr_ops.c b/net/sunrpc/xprtrdma/frwr_ops.c
index 524cac0a0715..0b6dad7580a1 100644
--- a/net/sunrpc/xprtrdma/frwr_ops.c
+++ b/net/sunrpc/xprtrdma/frwr_ops.c
@@ -542,7 +542,10 @@ static void frwr_wc_localinv_wake(struct ib_cq *cq, struct ib_wc *wc)
  * @req: rpcrdma_req with a non-empty list of MRs to process
  *
  * Sleeps until it is safe for the host CPU to access the previously mapped
- * memory regions.
+ * memory regions. This guarantees that registered MRs are properly fenced
+ * from the server before the RPC consumer accesses the data in them. It
+ * also ensures proper Send flow control: waking the next RPC waits until
+ * this RPC has relinquished all its Send Queue entries.
  */
 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
 {
@@ -616,3 +619,101 @@ void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
 		rpcrdma_mr_recycle(mr);
 	}
 }
+
+/**
+ * frwr_wc_localinv_done - Invoked by RDMA provider for a signaled LOCAL_INV WC
+ * @cq:	completion queue (ignored)
+ * @wc:	completed WR
+ *
+ */
+static void frwr_wc_localinv_done(struct ib_cq *cq, struct ib_wc *wc)
+{
+	struct ib_cqe *cqe = wc->wr_cqe;
+	struct rpcrdma_frwr *frwr =
+		container_of(cqe, struct rpcrdma_frwr, fr_cqe);
+	struct rpcrdma_mr *mr = container_of(frwr, struct rpcrdma_mr, frwr);
+
+	/* WARNING: Only wr_cqe and status are reliable at this point */
+	trace_xprtrdma_wc_li_done(wc, frwr);
+	rpcrdma_complete_rqst(frwr->fr_req->rl_reply);
+	__frwr_release_mr(wc, mr);
+}
+
+/**
+ * frwr_unmap_async - invalidate memory regions that were registered for @req
+ * @r_xprt: controlling transport instance
+ * @req: rpcrdma_req with a non-empty list of MRs to process
+ *
+ * This guarantees that registered MRs are properly fenced from the
+ * server before the RPC consumer accesses the data in them. It also
+ * ensures proper Send flow control: waking the next RPC waits until
+ * this RPC has relinquished all its Send Queue entries.
+ */
+void frwr_unmap_async(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
+{
+	struct ib_send_wr *first, *last, **prev;
+	const struct ib_send_wr *bad_wr;
+	struct rpcrdma_frwr *frwr;
+	struct rpcrdma_mr *mr;
+	int rc;
+
+	/* Chain the LOCAL_INV Work Requests and post them with
+	 * a single ib_post_send() call.
+	 */
+	frwr = NULL;
+	prev = &first;
+	while (!list_empty(&req->rl_registered)) {
+		mr = rpcrdma_mr_pop(&req->rl_registered);
+
+		trace_xprtrdma_mr_localinv(mr);
+		r_xprt->rx_stats.local_inv_needed++;
+
+		frwr = &mr->frwr;
+		frwr->fr_cqe.done = frwr_wc_localinv;
+		frwr->fr_req = req;
+		last = &frwr->fr_invwr;
+		last->next = NULL;
+		last->wr_cqe = &frwr->fr_cqe;
+		last->sg_list = NULL;
+		last->num_sge = 0;
+		last->opcode = IB_WR_LOCAL_INV;
+		last->send_flags = IB_SEND_SIGNALED;
+		last->ex.invalidate_rkey = mr->mr_handle;
+
+		*prev = last;
+		prev = &last->next;
+	}
+
+	/* Strong send queue ordering guarantees that when the
+	 * last WR in the chain completes, all WRs in the chain
+	 * are complete. The last completion will wake up the
+	 * RPC waiter.
+	 */
+	frwr->fr_cqe.done = frwr_wc_localinv_done;
+
+	/* Transport disconnect drains the receive CQ before it
+	 * replaces the QP. The RPC reply handler won't call us
+	 * unless ri_id->qp is a valid pointer.
+	 */
+	bad_wr = NULL;
+	rc = ib_post_send(r_xprt->rx_ia.ri_id->qp, first, &bad_wr);
+	trace_xprtrdma_post_send(req, rc);
+	if (!rc)
+		return;
+
+	/* Recycle MRs in the LOCAL_INV chain that did not get posted.
+	 */
+	while (bad_wr) {
+		frwr = container_of(bad_wr, struct rpcrdma_frwr, fr_invwr);
+		mr = container_of(frwr, struct rpcrdma_mr, frwr);
+		bad_wr = bad_wr->next;
+
+		rpcrdma_mr_recycle(mr);
+	}
+
+	/* The final LOCAL_INV WR in the chain is supposed to
+	 * do the wake. If it was never posted, the wake will
+	 * not happen, so wake here in that case.
+	 */
+	rpcrdma_complete_rqst(req->rl_reply);
+}