[PATCH v2 11/19] xprtrdma: Streamline rpcrdma_post_recvs

[Chuck Lever <chuck.lever@oracle.com>]

rb_lock is contended between rpcrdma_buffer_create,
rpcrdma_buffer_put, and rpcrdma_post_recvs.

Commit e340c2d6ef2a ("xprtrdma: Reduce the doorbell rate (Receive)")
causes rpcrdma_post_recvs to take the rb_lock repeatedly when it
determines more Receives are needed. Streamline this code path so
it takes the lock just once in most cases to build the Receive
chain that is about to be posted.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
---
 net/sunrpc/xprtrdma/verbs.c |   59 ++++++++++++++++++++++++++++---------------
 1 file changed, 38 insertions(+), 21 deletions(-)

diff --git a/net/sunrpc/xprtrdma/verbs.c b/net/sunrpc/xprtrdma/verbs.c
index de6be10..3270c8a 100644
--- a/net/sunrpc/xprtrdma/verbs.c
+++ b/net/sunrpc/xprtrdma/verbs.c
@@ -1478,11 +1478,13 @@ static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
 {
 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
-	struct ib_recv_wr *wr, *bad_wr;
+	struct ib_recv_wr *i, *wr, *bad_wr;
+	struct rpcrdma_rep *rep;
 	int needed, count, rc;
 
 	rc = 0;
 	count = 0;
+
 	needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
 	if (ep->rep_receive_count > needed)
 		goto out;
@@ -1490,39 +1492,48 @@ static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
 	if (!temp)
 		needed += RPCRDMA_MAX_RECV_BATCH;
 
-	count = 0;
+	/* fast path: all needed reps can be found on the free list */
 	wr = NULL;
+	spin_lock(&buf->rb_lock);
 	while (needed) {
-		struct rpcrdma_rep *rep;
-
-		spin_lock(&buf->rb_lock);
 		rep = list_first_entry_or_null(&buf->rb_recv_bufs,
 					       struct rpcrdma_rep, rr_list);
-		if (likely(rep))
-			list_del(&rep->rr_list);
-		spin_unlock(&buf->rb_lock);
-		if (!rep) {
-			rep = rpcrdma_rep_create(r_xprt, temp);
-			if (!rep)
-				break;
-		}
+		if (!rep)
+			break;
 
-		if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf)) {
-			rpcrdma_recv_buffer_put(rep);
+		list_del(&rep->rr_list);
+		rep->rr_recv_wr.next = wr;
+		wr = &rep->rr_recv_wr;
+		--needed;
+	}
+	spin_unlock(&buf->rb_lock);
+
+	while (needed) {
+		rep = rpcrdma_rep_create(r_xprt, temp);
+		if (!rep)
 			break;
-		}
 
-		trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
 		rep->rr_recv_wr.next = wr;
 		wr = &rep->rr_recv_wr;
-		++count;
 		--needed;
 	}
-	if (!count)
+	if (!wr)
 		goto out;
 
+	for (i = wr; i; i = i->next) {
+		rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
+
+		if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
+			goto release_wrs;
+
+		trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
+		++count;
+	}
+
 	rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
 			  (const struct ib_recv_wr **)&bad_wr);
+out:
+	trace_xprtrdma_post_recvs(r_xprt, count, rc);
 	if (rc) {
 		for (wr = bad_wr; wr;) {
 			struct rpcrdma_rep *rep;
@@ -1534,6 +1545,12 @@ static void rpcrdma_regbuf_free(struct rpcrdma_regbuf *rb)
 		}
 	}
 	ep->rep_receive_count += count;
-out:
-	trace_xprtrdma_post_recvs(r_xprt, count, rc);
+	return;
+
+release_wrs:
+	for (i = wr; i;) {
+		rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
+		i = i->next;
+		rpcrdma_recv_buffer_put(rep);
+	}
 }