[PATCH net-next v2 0/2] net/smc: send and write inline optimization for smc

Send cdc msgs and write data inline if qp has sufficent inline
space, helps latency reducing. 

In my test environment, which are 2 VMs running on the same
physical host and whose NICs(ConnectX-4Lx) are working on
SR-IOV mode, qperf shows 0.4us-1.3us improvement in latency.

Test command:
server: smc_run taskset -c 1 qperf
client: smc_run taskset -c 1 qperf <server ip> -oo \
		msg_size:1:2K:*2 -t 30 -vu tcp_lat

The results shown below:
msgsize     before       after
1B          11.9 us      10.6 us (-1.3 us)
2B          11.7 us      10.7 us (-1.0 us)
4B          11.7 us      10.7 us (-1.0 us)
8B          11.6 us      10.6 us (-1.0 us)
16B         11.7 us      10.7 us (-1.0 us)
32B         11.7 us      10.6 us (-1.1 us)
64B         11.7 us      11.2 us (-0.5 us)
128B        11.6 us      11.2 us (-0.4 us)
256B        11.8 us      11.2 us (-0.6 us)
512B        11.8 us      11.3 us (-0.5 us)
1KB         11.9 us      11.5 us (-0.4 us)
2KB         12.1 us      11.5 us (-0.6 us)

Guangguan Wang (2):
  net/smc: send cdc msg inline if qp has sufficient inline space
  net/smc: rdma write inline if qp has sufficient inline space

 net/smc/smc_ib.c |  1 +
 net/smc/smc_tx.c | 17 ++++++++++++-----
 net/smc/smc_wr.c |  5 ++++-
 3 files changed, 17 insertions(+), 6 deletions(-)

-- 
2.24.3 (Apple Git-128)

Rdma write with inline flag when sending small packages,
whose length is shorter than the qp's max_inline_data, can
help reducing latency.

In my test environment, which are 2 VMs running on the same
physical host and whose NICs(ConnectX-4Lx) are working on
SR-IOV mode, qperf shows 0.5us-0.7us improvement in latency.

Test command:
server: smc_run taskset -c 1 qperf
client: smc_run taskset -c 1 qperf <server ip> -oo \
		msg_size:1:2K:*2 -t 30 -vu tcp_lat

The results shown below:
msgsize     before       after
1B          11.2 us      10.6 us (-0.6 us)
2B          11.2 us      10.7 us (-0.5 us)
4B          11.3 us      10.7 us (-0.6 us)
8B          11.2 us      10.6 us (-0.6 us)
16B         11.3 us      10.7 us (-0.6 us)
32B         11.3 us      10.6 us (-0.7 us)
64B         11.2 us      11.2 us (0 us)
128B        11.2 us      11.2 us (0 us)
256B        11.2 us      11.2 us (0 us)
512B        11.4 us      11.3 us (-0.1 us)
1KB         11.4 us      11.5 us (0.1 us)
2KB         11.5 us      11.5 us (0 us)

Reported-by: kernel test robot <redacted>
Signed-off-by: Guangguan Wang <redacted>
---
 net/smc/smc_tx.c | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

diff --git a/net/smc/smc_tx.c b/net/smc/smc_tx.c
index 98ca9229fe87..805a546e8c04 100644
--- a/net/smc/smc_tx.c
+++ b/net/smc/smc_tx.c

@@ -391,12 +391,20 @@ static int smcr_tx_rdma_writes(struct smc_connection *conn, size_t len,
 	int rc;
 
 	for (dstchunk = 0; dstchunk < 2; dstchunk++) {
-		struct ib_sge *sge =
-			wr_rdma_buf->wr_tx_rdma[dstchunk].wr.sg_list;
+		struct ib_rdma_wr *wr = &wr_rdma_buf->wr_tx_rdma[dstchunk];
+		struct ib_sge *sge = wr->wr.sg_list;
+		u64 base_addr = dma_addr;
+
+		if (dst_len < link->qp_attr.cap.max_inline_data) {
+			base_addr = (uintptr_t)conn->sndbuf_desc->cpu_addr;
+			wr->wr.send_flags |= IB_SEND_INLINE;
+		} else {
+			wr->wr.send_flags &= ~IB_SEND_INLINE;
+		}
 
 		num_sges = 0;
 		for (srcchunk = 0; srcchunk < 2; srcchunk++) {
-			sge[srcchunk].addr = dma_addr + src_off;
+			sge[srcchunk].addr = base_addr + src_off;
 			sge[srcchunk].length = src_len;
 			num_sges++;
 

@@ -410,8 +418,7 @@ static int smcr_tx_rdma_writes(struct smc_connection *conn, size_t len,
 			src_len = dst_len - src_len; /* remainder */
 			src_len_sum += src_len;
 		}
-		rc = smc_tx_rdma_write(conn, dst_off, num_sges,
-				       &wr_rdma_buf->wr_tx_rdma[dstchunk]);
+		rc = smc_tx_rdma_write(conn, dst_off, num_sges, wr);
 		if (rc)
 			return rc;
 		if (dst_len_sum == len)

-- 
2.24.3 (Apple Git-128)

As cdc msg's length is 44B, cdc msgs can be sent inline in
most rdma devices, which can help reducing sending latency.

In my test environment, which are 2 VMs running on the same
physical host and whose NICs(ConnectX-4Lx) are working on
SR-IOV mode, qperf shows 0.4us-0.7us improvement in latency.

Test command:
server: smc_run taskset -c 1 qperf
client: smc_run taskset -c 1 qperf <server ip> -oo \
		msg_size:1:2K:*2 -t 30 -vu tcp_lat

The results shown below:
msgsize     before       after
1B          11.9 us      11.2 us (-0.7 us)
2B          11.7 us      11.2 us (-0.5 us)
4B          11.7 us      11.3 us (-0.4 us)
8B          11.6 us      11.2 us (-0.4 us)
16B         11.7 us      11.3 us (-0.4 us)
32B         11.7 us      11.3 us (-0.4 us)
64B         11.7 us      11.2 us (-0.5 us)
128B        11.6 us      11.2 us (-0.4 us)
256B        11.8 us      11.2 us (-0.6 us)
512B        11.8 us      11.4 us (-0.4 us)
1KB         11.9 us      11.4 us (-0.5 us)
2KB         12.1 us      11.5 us (-0.6 us)

Reported-by: kernel test robot <redacted>
Signed-off-by: Guangguan Wang <redacted>
---
 net/smc/smc_ib.c | 1 +
 net/smc/smc_wr.c | 5 ++++-
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/net/smc/smc_ib.c b/net/smc/smc_ib.c
index a3e2d3b89568..dcda4165d107 100644
--- a/net/smc/smc_ib.c
+++ b/net/smc/smc_ib.c

@@ -671,6 +671,7 @@ int smc_ib_create_queue_pair(struct smc_link *lnk)
 			.max_recv_wr = SMC_WR_BUF_CNT * 3,
 			.max_send_sge = SMC_IB_MAX_SEND_SGE,
 			.max_recv_sge = sges_per_buf,
+			.max_inline_data = 0,
 		},
 		.sq_sig_type = IB_SIGNAL_REQ_WR,
 		.qp_type = IB_QPT_RC,

diff --git a/net/smc/smc_wr.c b/net/smc/smc_wr.c
index 24be1d03fef9..26f8f240d9e8 100644
--- a/net/smc/smc_wr.c
+++ b/net/smc/smc_wr.c

@@ -554,10 +554,11 @@ void smc_wr_remember_qp_attr(struct smc_link *lnk)
 static void smc_wr_init_sge(struct smc_link *lnk)
 {
 	int sges_per_buf = (lnk->lgr->smc_version == SMC_V2) ? 2 : 1;
+	bool send_inline = (lnk->qp_attr.cap.max_inline_data > SMC_WR_TX_SIZE);
 	u32 i;
 
 	for (i = 0; i < lnk->wr_tx_cnt; i++) {
-		lnk->wr_tx_sges[i].addr =
+		lnk->wr_tx_sges[i].addr = send_inline ? (uintptr_t)(&lnk->wr_tx_bufs[i]) :
 			lnk->wr_tx_dma_addr + i * SMC_WR_BUF_SIZE;
 		lnk->wr_tx_sges[i].length = SMC_WR_TX_SIZE;
 		lnk->wr_tx_sges[i].lkey = lnk->roce_pd->local_dma_lkey;

@@ -575,6 +576,8 @@ static void smc_wr_init_sge(struct smc_link *lnk)
 		lnk->wr_tx_ibs[i].opcode = IB_WR_SEND;
 		lnk->wr_tx_ibs[i].send_flags =
 			IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+		if (send_inline)
+			lnk->wr_tx_ibs[i].send_flags |= IB_SEND_INLINE;
 		lnk->wr_tx_rdmas[i].wr_tx_rdma[0].wr.opcode = IB_WR_RDMA_WRITE;
 		lnk->wr_tx_rdmas[i].wr_tx_rdma[1].wr.opcode = IB_WR_RDMA_WRITE;
 		lnk->wr_tx_rdmas[i].wr_tx_rdma[0].wr.sg_list =

-- 
2.24.3 (Apple Git-128)

On Sat, May 14, 2022 at 06:27:38PM +0800, Guangguan Wang wrote:

As cdc msg's length is 44B, cdc msgs can be sent inline in
most rdma devices, which can help reducing sending latency.

In my test environment, which are 2 VMs running on the same
physical host and whose NICs(ConnectX-4Lx) are working on
SR-IOV mode, qperf shows 0.4us-0.7us improvement in latency.

Test command:
server: smc_run taskset -c 1 qperf
client: smc_run taskset -c 1 qperf <server ip> -oo \
		msg_size:1:2K:*2 -t 30 -vu tcp_lat

The results shown below:
msgsize     before       after
1B          11.9 us      11.2 us (-0.7 us)
2B          11.7 us      11.2 us (-0.5 us)
4B          11.7 us      11.3 us (-0.4 us)
8B          11.6 us      11.2 us (-0.4 us)
16B         11.7 us      11.3 us (-0.4 us)
32B         11.7 us      11.3 us (-0.4 us)
64B         11.7 us      11.2 us (-0.5 us)
128B        11.6 us      11.2 us (-0.4 us)
256B        11.8 us      11.2 us (-0.6 us)
512B        11.8 us      11.4 us (-0.4 us)
1KB         11.9 us      11.4 us (-0.5 us)
2KB         12.1 us      11.5 us (-0.6 us)

Reported-by: kernel test robot <redacted>

You don't need to add this tag, this tag represents who found the issue.
Tested-by is reasonable.

Signed-off-by: Guangguan Wang <redacted>

Reviewed-by: Tony Lu <tonylu@linux.alibaba.com>

Thanks,
Tony Lu