summaryrefslogtreecommitdiffstats
path: root/net/sunrpc
diff options
context:
space:
mode:
authorChuck Lever <chuck.lever@oracle.com>2019-10-17 14:31:53 -0400
committerAnna Schumaker <Anna.Schumaker@Netapp.com>2019-10-24 10:30:40 -0400
commit614f3c96d7e5efd1c4dc699524857130a52c6a7f (patch)
tree05b6a845cd5731657b7f63c5af28a597b2bedd38 /net/sunrpc
parentd6764bbd7763fa9d669bba7fc5a50a4bdd8f591b (diff)
downloadlinux-614f3c96d7e5efd1c4dc699524857130a52c6a7f.tar.gz
linux-614f3c96d7e5efd1c4dc699524857130a52c6a7f.tar.bz2
linux-614f3c96d7e5efd1c4dc699524857130a52c6a7f.zip
xprtrdma: Pull up sometimes
On some platforms, DMA mapping part of a page is more costly than copying bytes. Restore the pull-up code and use that when we think it's going to be faster. The heuristic for now is to pull-up when the size of the RPC message body fits in the buffer underlying the head iovec. 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, as is handling a Send completion, 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. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Diffstat (limited to 'net/sunrpc')
-rw-r--r--net/sunrpc/xprtrdma/backchannel.c2
-rw-r--r--net/sunrpc/xprtrdma/rpc_rdma.c82
-rw-r--r--net/sunrpc/xprtrdma/verbs.c2
-rw-r--r--net/sunrpc/xprtrdma/xprt_rdma.h2
4 files changed, 81 insertions, 7 deletions
diff --git a/net/sunrpc/xprtrdma/backchannel.c b/net/sunrpc/xprtrdma/backchannel.c
index 50e075fcdd8f..11685245546a 100644
--- a/net/sunrpc/xprtrdma/backchannel.c
+++ b/net/sunrpc/xprtrdma/backchannel.c
@@ -79,7 +79,7 @@ static int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)
*p = xdr_zero;
if (rpcrdma_prepare_send_sges(r_xprt, req, RPCRDMA_HDRLEN_MIN,
- &rqst->rq_snd_buf, rpcrdma_noch))
+ &rqst->rq_snd_buf, rpcrdma_noch_pullup))
return -EIO;
trace_xprtrdma_cb_reply(rqst);
diff --git a/net/sunrpc/xprtrdma/rpc_rdma.c b/net/sunrpc/xprtrdma/rpc_rdma.c
index a441dbf9f198..4ad88893e964 100644
--- a/net/sunrpc/xprtrdma/rpc_rdma.c
+++ b/net/sunrpc/xprtrdma/rpc_rdma.c
@@ -392,7 +392,7 @@ static int rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt,
unsigned int pos;
int nsegs;
- if (rtype == rpcrdma_noch)
+ if (rtype == rpcrdma_noch_pullup || rtype == rpcrdma_noch_mapped)
goto done;
pos = rqst->rq_snd_buf.head[0].iov_len;
@@ -691,6 +691,72 @@ out_mapping_err:
return false;
}
+/* Copy the tail to the end of the head buffer.
+ */
+static void rpcrdma_pullup_tail_iov(struct rpcrdma_xprt *r_xprt,
+ struct rpcrdma_req *req,
+ struct xdr_buf *xdr)
+{
+ unsigned char *dst;
+
+ dst = (unsigned char *)xdr->head[0].iov_base;
+ dst += xdr->head[0].iov_len + xdr->page_len;
+ memmove(dst, xdr->tail[0].iov_base, xdr->tail[0].iov_len);
+ r_xprt->rx_stats.pullup_copy_count += xdr->tail[0].iov_len;
+}
+
+/* Copy pagelist content into the head buffer.
+ */
+static void rpcrdma_pullup_pagelist(struct rpcrdma_xprt *r_xprt,
+ struct rpcrdma_req *req,
+ struct xdr_buf *xdr)
+{
+ unsigned int len, page_base, remaining;
+ struct page **ppages;
+ unsigned char *src, *dst;
+
+ dst = (unsigned char *)xdr->head[0].iov_base;
+ dst += xdr->head[0].iov_len;
+ ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT);
+ page_base = offset_in_page(xdr->page_base);
+ remaining = xdr->page_len;
+ while (remaining) {
+ src = page_address(*ppages);
+ src += page_base;
+ len = min_t(unsigned int, PAGE_SIZE - page_base, remaining);
+ memcpy(dst, src, len);
+ r_xprt->rx_stats.pullup_copy_count += len;
+
+ ppages++;
+ dst += len;
+ remaining -= len;
+ page_base = 0;
+ }
+}
+
+/* Copy the contents of @xdr into @rl_sendbuf and DMA sync it.
+ * When the head, pagelist, and tail are small, a pull-up copy
+ * is considerably less costly than DMA mapping the components
+ * of @xdr.
+ *
+ * Assumptions:
+ * - the caller has already verified that the total length
+ * of the RPC Call body will fit into @rl_sendbuf.
+ */
+static bool rpcrdma_prepare_noch_pullup(struct rpcrdma_xprt *r_xprt,
+ struct rpcrdma_req *req,
+ struct xdr_buf *xdr)
+{
+ if (unlikely(xdr->tail[0].iov_len))
+ rpcrdma_pullup_tail_iov(r_xprt, req, xdr);
+
+ if (unlikely(xdr->page_len))
+ rpcrdma_pullup_pagelist(r_xprt, req, xdr);
+
+ /* The whole RPC message resides in the head iovec now */
+ return rpcrdma_prepare_head_iov(r_xprt, req, xdr->len);
+}
+
static bool rpcrdma_prepare_noch_mapped(struct rpcrdma_xprt *r_xprt,
struct rpcrdma_req *req,
struct xdr_buf *xdr)
@@ -779,7 +845,11 @@ inline int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
goto out_unmap;
switch (rtype) {
- case rpcrdma_noch:
+ case rpcrdma_noch_pullup:
+ if (!rpcrdma_prepare_noch_pullup(r_xprt, req, xdr))
+ goto out_unmap;
+ break;
+ case rpcrdma_noch_mapped:
if (!rpcrdma_prepare_noch_mapped(r_xprt, req, xdr))
goto out_unmap;
break;
@@ -827,6 +897,7 @@ rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
struct xdr_stream *xdr = &req->rl_stream;
enum rpcrdma_chunktype rtype, wtype;
+ struct xdr_buf *buf = &rqst->rq_snd_buf;
bool ddp_allowed;
__be32 *p;
int ret;
@@ -884,8 +955,9 @@ rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
*/
if (rpcrdma_args_inline(r_xprt, rqst)) {
*p++ = rdma_msg;
- rtype = rpcrdma_noch;
- } else if (ddp_allowed && rqst->rq_snd_buf.flags & XDRBUF_WRITE) {
+ rtype = buf->len < rdmab_length(req->rl_sendbuf) ?
+ rpcrdma_noch_pullup : rpcrdma_noch_mapped;
+ } else if (ddp_allowed && buf->flags & XDRBUF_WRITE) {
*p++ = rdma_msg;
rtype = rpcrdma_readch;
} else {
@@ -927,7 +999,7 @@ rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst)
goto out_err;
ret = rpcrdma_prepare_send_sges(r_xprt, req, req->rl_hdrbuf.len,
- &rqst->rq_snd_buf, rtype);
+ buf, rtype);
if (ret)
goto out_err;
diff --git a/net/sunrpc/xprtrdma/verbs.c b/net/sunrpc/xprtrdma/verbs.c
index 2f4658255343..a514e2c89ac3 100644
--- a/net/sunrpc/xprtrdma/verbs.c
+++ b/net/sunrpc/xprtrdma/verbs.c
@@ -1165,7 +1165,7 @@ int rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
for (i = 0; i < buf->rb_max_requests; i++) {
struct rpcrdma_req *req;
- req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE,
+ req = rpcrdma_req_create(r_xprt, RPCRDMA_V1_DEF_INLINE_SIZE * 2,
GFP_KERNEL);
if (!req)
goto out;
diff --git a/net/sunrpc/xprtrdma/xprt_rdma.h b/net/sunrpc/xprtrdma/xprt_rdma.h
index cdd6a3d43e0f..5d15140a0266 100644
--- a/net/sunrpc/xprtrdma/xprt_rdma.h
+++ b/net/sunrpc/xprtrdma/xprt_rdma.h
@@ -554,6 +554,8 @@ void frwr_unmap_async(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
enum rpcrdma_chunktype {
rpcrdma_noch = 0,
+ rpcrdma_noch_pullup,
+ rpcrdma_noch_mapped,
rpcrdma_readch,
rpcrdma_areadch,
rpcrdma_writech,