svcrdma: Avoid DMA mapping small RPC Replies

On some platforms, DMA mapping part of a page is more costly than
copying bytes. Indeed, not involving the I/O MMU can help the
RPC/RDMA transport scale better for tiny I/Os across more RDMA
devices. This is because interaction with the I/O MMU is eliminated
for each of these small I/Os. Without the explicit unmapping, the
NIC no longer needs to do a costly internal TLB shoot down for
buffers that are just a handful of bytes.

Since pull-up is now a more a frequent operation, I've introduced a
trace point in the pull-up path. It can be used for debugging or
user-space tools that count pull-up frequency.

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

include/linux/sunrpc/svc_rdma.h

@@ -52,6 +52,7 @@
 
 /* Default and maximum inline threshold sizes */
 enum {
+	RPCRDMA_PULLUP_THRESH = RPCRDMA_V1_DEF_INLINE_SIZE >> 1,
 	RPCRDMA_DEF_INLINE_THRESH = 4096,
 	RPCRDMA_MAX_INLINE_THRESH = 65536
 };

include/trace/events/rpcrdma.h

@@ -1639,6 +1639,24 @@ TRACE_EVENT(svcrdma_dma_map_rwctx,
 	)
 );
 
+TRACE_EVENT(svcrdma_send_pullup,
+	TP_PROTO(
+		unsigned int len
+	),
+
+	TP_ARGS(len),
+
+	TP_STRUCT__entry(
+		__field(unsigned int, len)
+	),
+
+	TP_fast_assign(
+		__entry->len = len;
+	),
+
+	TP_printk("len=%u", __entry->len)
+);
+
 TRACE_EVENT(svcrdma_send_failed,
 	TP_PROTO(
 		const struct svc_rqst *rqst,

net/sunrpc/xprtrdma/svc_rdma_sendto.c

@@ -541,6 +541,7 @@ static int svc_rdma_dma_map_buf(struct svcxprt_rdma *rdma,
 /**
  * svc_rdma_pull_up_needed - Determine whether to use pull-up
  * @rdma: controlling transport
+ * @sctxt: send_ctxt for the Send WR
  * @rctxt: Write and Reply chunks provided by client
  * @xdr: xdr_buf containing RPC message to transmit
  *
@@ -549,11 +550,20 @@ static int svc_rdma_dma_map_buf(struct svcxprt_rdma *rdma,
  *	%false otherwise
  */
 static bool svc_rdma_pull_up_needed(struct svcxprt_rdma *rdma,
+				    struct svc_rdma_send_ctxt *sctxt,
 				    const struct svc_rdma_recv_ctxt *rctxt,
 				    struct xdr_buf *xdr)
 {
 	int elements;
 
+	/* For small messages, copying bytes is cheaper than DMA mapping.
+	 */
+	if (sctxt->sc_hdrbuf.len + xdr->len < RPCRDMA_PULLUP_THRESH)
+		return true;
+
+	/* Check whether the xdr_buf has more elements than can
+	 * fit in a single RDMA Send.
+	 */
 	/* xdr->head */
 	elements = 1;
 
@@ -636,6 +646,7 @@ static int svc_rdma_pull_up_reply_msg(struct svcxprt_rdma *rdma,
 		memcpy(dst, tailbase, taillen);
 
 	sctxt->sc_sges[0].length += xdr->len;
+	trace_svcrdma_send_pullup(sctxt->sc_sges[0].length);
 	return 0;
 }
 
@@ -675,7 +686,7 @@ int svc_rdma_map_reply_msg(struct svcxprt_rdma *rdma,
 	/* For pull-up, svc_rdma_send() will sync the transport header.
 	 * No additional DMA mapping is necessary.
 	 */
-	if (svc_rdma_pull_up_needed(rdma, rctxt, xdr))
+	if (svc_rdma_pull_up_needed(rdma, sctxt, rctxt, xdr))
 		return svc_rdma_pull_up_reply_msg(rdma, sctxt, rctxt, xdr);
 
 	++sctxt->sc_cur_sge_no;