From: Chuck Lever <chuck.lever@xxxxxxxxxx>
I've noticed that using the ib-comp-wq workqueue delays Send
Completions anywhere between 5us and 3 or more milliseconds.
For RDMA Write and Send completions, this is not a terribly
significant issue, since these just release resources. They do not
contribute to RPC round-trip time.
However, for RDMA Read completions, it delays the start of NFS
WRITE operations, adding round-trip latency.
For small to moderate NFS WRITEs, using soft IRQ completion means
up to 5us better latency per NFS WRITE -- this is a significant
portion of average RTT for small NFS WRITEs, which is 40-75us.
Signed-off-by: Chuck Lever <chuck.lever@xxxxxxxxxx>
---
net/sunrpc/xprtrdma/svc_rdma_rw.c | 4 ++--
net/sunrpc/xprtrdma/svc_rdma_sendto.c | 6 +++---
net/sunrpc/xprtrdma/svc_rdma_transport.c | 2 +-
3 files changed, 6 insertions(+), 6 deletions(-)
diff --git a/net/sunrpc/xprtrdma/svc_rdma_rw.c b/net/sunrpc/xprtrdma/svc_rdma_rw.c
index e460e25a1d6d..ada164c027bc 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_rw.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_rw.c
@@ -56,9 +56,9 @@ svc_rdma_get_rw_ctxt(struct svcxprt_rdma *rdma, unsigned int sges)
struct svc_rdma_rw_ctxt *ctxt;
struct llist_node *node;
- spin_lock(&rdma->sc_rw_ctxt_lock);
+ spin_lock_bh(&rdma->sc_rw_ctxt_lock);
node = llist_del_first(&rdma->sc_rw_ctxts);
- spin_unlock(&rdma->sc_rw_ctxt_lock);
+ spin_unlock_bh(&rdma->sc_rw_ctxt_lock);
if (node) {
ctxt = llist_entry(node, struct svc_rdma_rw_ctxt, rw_node);
} else {
diff --git a/net/sunrpc/xprtrdma/svc_rdma_sendto.c b/net/sunrpc/xprtrdma/svc_rdma_sendto.c
index e27345af6289..49a9f409bc8e 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_sendto.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_sendto.c
@@ -198,12 +198,13 @@ struct svc_rdma_send_ctxt *svc_rdma_send_ctxt_get(struct svcxprt_rdma *rdma)
struct svc_rdma_send_ctxt *ctxt;
struct llist_node *node;
- spin_lock(&rdma->sc_send_lock);
+ spin_lock_bh(&rdma->sc_send_lock);
node = llist_del_first(&rdma->sc_send_ctxts);
+ spin_unlock_bh(&rdma->sc_send_lock);
if (!node)
goto out_empty;
+
ctxt = llist_entry(node, struct svc_rdma_send_ctxt, sc_node);
- spin_unlock(&rdma->sc_send_lock);
out:
rpcrdma_set_xdrlen(&ctxt->sc_hdrbuf, 0);
@@ -216,7 +217,6 @@ struct svc_rdma_send_ctxt *svc_rdma_send_ctxt_get(struct svcxprt_rdma *rdma)
return ctxt;
out_empty:
- spin_unlock(&rdma->sc_send_lock);
ctxt = svc_rdma_send_ctxt_alloc(rdma);
if (!ctxt)
return NULL;
diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
index 7bd50efeeb4e..8de32927cd7d 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_transport.c
+++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
@@ -430,7 +430,7 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
goto errout;
}
newxprt->sc_sq_cq = ib_alloc_cq_any(dev, newxprt, newxprt->sc_sq_depth,
- IB_POLL_WORKQUEUE);
+ IB_POLL_SOFTIRQ);
if (IS_ERR(newxprt->sc_sq_cq))
goto errout;
newxprt->sc_rq_cq = ib_alloc_cq_any(dev, newxprt, rq_depth,
