summaryrefslogtreecommitdiff
path: root/net/sunrpc/xprtrdma
diff options
context:
space:
mode:
authorChuck Lever <chuck.lever@oracle.com>2020-03-03 21:28:14 +0300
committerChuck Lever <chuck.lever@oracle.com>2020-03-16 19:04:33 +0300
commit0dabe948f28274e7956a625a24f205016b810693 (patch)
tree520f4aa1ad1e6973d3e5d0914927988134548644 /net/sunrpc/xprtrdma
parentaee4b74a3f273b54d136132fedf575ec464f4134 (diff)
downloadlinux-0dabe948f28274e7956a625a24f205016b810693.tar.xz
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>
Diffstat (limited to 'net/sunrpc/xprtrdma')
-rw-r--r--net/sunrpc/xprtrdma/svc_rdma_sendto.c13
1 files changed, 12 insertions, 1 deletions
diff --git a/net/sunrpc/xprtrdma/svc_rdma_sendto.c b/net/sunrpc/xprtrdma/svc_rdma_sendto.c
index 7b9853214769..90cba3058f04 100644
--- a/net/sunrpc/xprtrdma/svc_rdma_sendto.c
+++ b/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;