[PATCH net-next 0/7] net/smc: some datapath performance optimizations

Hi:

This series tries to improve the performance of SMC in datapath.

- patch #1, add sysctl interface to support tuning the behaviour of
  SMC in container environment.

- patch #2/#3, add autocorking support which is very efficient for small
  messages without trade-off for latency.

- patch #4, send directly on setting TCP_NODELAY, without wake up the
  TX worker, this make it consistent with clearing TCP_CORK.

- patch #5, this correct the setting of RMB window update limit, so
  we don't send CDC messages to update peer's RMB window too frequently
  in some cases.

- patch #6, implemented something like NAPI in SMC, decrease the number
  of hardirq when busy.

- patch #7, this moves TX work doing in the BH to the user context when
  sock_lock is hold by user.


With this patchset applied, we can get a good performance gain:
- qperf tcp_bw test has shown a great improvement. Other benchmarks like
  'netperf TCP_STREAM' or 'sockperf throughput' has similar result.
- In my testing environment, running qperf tcp_bw and tcp_lat, SMC behaves
  better then TCP in most all message size.

Here are some test results with the following testing command:
client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \
		-t 30 -vu tcp_{bw|lat}
server: smc_run taskset -c 1 qperf

==== Bandwidth ====
 MsgSize        Origin SMC              TCP                SMC with patches
       1         0.578 MB/s      2.392 MB/s(313.57%)      2.561 MB/s(342.83%)
       2         1.159 MB/s      4.780 MB/s(312.53%)      5.162 MB/s(345.46%)
       4         2.283 MB/s     10.266 MB/s(349.77%)     10.122 MB/s(343.46%)
       8         4.668 MB/s     19.040 MB/s(307.86%)     20.521 MB/s(339.59%)
      16         9.147 MB/s     38.904 MB/s(325.31%)     40.823 MB/s(346.29%)
      32        18.369 MB/s     79.587 MB/s(333.25%)     80.535 MB/s(338.42%)
      64        36.562 MB/s    148.668 MB/s(306.61%)    158.170 MB/s(332.60%)
     128        72.961 MB/s    274.913 MB/s(276.80%)    316.217 MB/s(333.41%)
     256       144.705 MB/s    512.059 MB/s(253.86%)    626.019 MB/s(332.62%)
     512       288.873 MB/s    884.977 MB/s(206.35%)   1221.596 MB/s(322.88%)
    1024       574.180 MB/s   1337.736 MB/s(132.98%)   2203.156 MB/s(283.70%)
    2048      1095.192 MB/s   1865.952 MB/s( 70.38%)   3036.448 MB/s(177.25%)
    4096      2066.157 MB/s   2380.337 MB/s( 15.21%)   3834.271 MB/s( 85.58%)
    8192      3717.198 MB/s   2733.073 MB/s(-26.47%)   4904.910 MB/s( 31.95%)
   16384      4742.221 MB/s   2958.693 MB/s(-37.61%)   5220.272 MB/s( 10.08%)
   32768      5349.550 MB/s   3061.285 MB/s(-42.77%)   5321.865 MB/s( -0.52%)
   65536      5162.919 MB/s   3731.408 MB/s(-27.73%)   5245.021 MB/s(  1.59%)
==== Latency ====
 MsgSize        Origin SMC              TCP                SMC with patches
       1        10.540 us     11.938 us( 13.26%)         10.356 us( -1.75%)
       2        10.996 us     11.992 us(  9.06%)         10.073 us( -8.39%)
       4        10.229 us     11.687 us( 14.25%)          9.996 us( -2.28%)
       8        10.203 us     11.653 us( 14.21%)         10.063 us( -1.37%)
      16        10.530 us     11.313 us(  7.44%)         10.013 us( -4.91%)
      32        10.241 us     11.586 us( 13.13%)         10.081 us( -1.56%)
      64        10.693 us     11.652 us(  8.97%)          9.986 us( -6.61%)
     128        10.597 us     11.579 us(  9.27%)         10.262 us( -3.16%)
     256        10.409 us     11.957 us( 14.87%)         10.148 us( -2.51%)
     512        11.088 us     12.505 us( 12.78%)         10.206 us( -7.95%)
    1024        11.240 us     12.255 us(  9.03%)         10.631 us( -5.42%)
    2048        11.485 us     16.970 us( 47.76%)         10.981 us( -4.39%)
    4096        12.077 us     13.948 us( 15.49%)         11.847 us( -1.90%)
    8192        13.683 us     16.693 us( 22.00%)         13.336 us( -2.54%)
   16384        16.470 us     23.615 us( 43.38%)         16.519 us(  0.30%)
   32768        22.540 us     40.966 us( 81.75%)         22.452 us( -0.39%)
   65536        34.192 us     73.003 us(113.51%)         33.916 us( -0.81%)

------------
Test environment notes:
1. Testing is run on 2 VMs within the same physical host
2. The NIC is ConnectX-4Lx, using SRIOV, and passing through 2 VFs to the
   2 VMs respectively.
3. To decrease jitter, VM's vCPU are binded to each physical CPU, and those
   physical CPUs are all isolated using boot parameter `isolcpus=xxx`
4. The queue number are set to 1, and interrupt from the queue is binded to
   CPU0 in the guest


Dust Li (7):
  net/smc: add sysctl interface for SMC
  net/smc: add autocorking support
  net/smc: add sysctl for autocorking
  net/smc: send directly on setting TCP_NODELAY
  net/smc: correct settings of RMB window update limit
  net/smc: don't req_notify until all CQEs drained
  net/smc: don't send in the BH context if sock_owned_by_user

 Documentation/networking/smc-sysctl.rst |  23 +++++
 include/net/netns/smc.h                 |   4 +
 net/smc/Makefile                        |   2 +-
 net/smc/af_smc.c                        |  30 ++++++-
 net/smc/smc.h                           |   6 ++
 net/smc/smc_cdc.c                       |  24 ++++--
 net/smc/smc_core.c                      |   2 +-
 net/smc/smc_sysctl.c                    |  80 ++++++++++++++++++
 net/smc/smc_sysctl.h                    |  32 +++++++
 net/smc/smc_tx.c                        | 107 +++++++++++++++++++++---
 net/smc/smc_wr.c                        |  49 ++++++-----
 11 files changed, 317 insertions(+), 42 deletions(-)
 create mode 100644 Documentation/networking/smc-sysctl.rst
 create mode 100644 net/smc/smc_sysctl.c
 create mode 100644 net/smc/smc_sysctl.h

-- 
2.19.1.3.ge56e4f7

This patch add sysctl interface to support container environment
for SMC as we talk in the mail list.

Link: https://lore.kernel.org/netdev/20220224020253.GF5443@linux.alibaba.com
Co-developed-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
---
 include/net/netns/smc.h |  3 ++
 net/smc/Makefile        |  2 +-
 net/smc/af_smc.c        | 10 ++++++
 net/smc/smc_sysctl.c    | 70 +++++++++++++++++++++++++++++++++++++++++
 net/smc/smc_sysctl.h    | 32 +++++++++++++++++++
 5 files changed, 116 insertions(+), 1 deletion(-)
 create mode 100644 net/smc/smc_sysctl.c
 create mode 100644 net/smc/smc_sysctl.h

diff --git a/include/net/netns/smc.h b/include/net/netns/smc.h
index 47b166684fd8..1682eae50579 100644
--- a/include/net/netns/smc.h
+++ b/include/net/netns/smc.h

@@ -14,5 +14,8 @@ struct netns_smc {
 	struct smc_stats_rsn		*fback_rsn;
 
 	bool				limit_smc_hs;	/* constraint on handshake */
+#ifdef CONFIG_SYSCTL
+	struct ctl_table_header		*smc_hdr;
+#endif
 };
 #endif

diff --git a/net/smc/Makefile b/net/smc/Makefile
index 196fb6f01b14..640af9a39f9c 100644
--- a/net/smc/Makefile
+++ b/net/smc/Makefile

@@ -4,4 +4,4 @@ obj-$(CONFIG_SMC)	+= smc.o
 obj-$(CONFIG_SMC_DIAG)	+= smc_diag.o
 smc-y := af_smc.o smc_pnet.o smc_ib.o smc_clc.o smc_core.o smc_wr.o smc_llc.o
 smc-y += smc_cdc.o smc_tx.o smc_rx.o smc_close.o smc_ism.o smc_netlink.o smc_stats.o
-smc-y += smc_tracepoint.o
+smc-y += smc_tracepoint.o smc_sysctl.o

diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
index 30acc31b2c45..19b3066cf7af 100644
--- a/net/smc/af_smc.c
+++ b/net/smc/af_smc.c

@@ -51,6 +51,7 @@
 #include "smc_close.h"
 #include "smc_stats.h"
 #include "smc_tracepoint.h"
+#include "smc_sysctl.h"
 
 static DEFINE_MUTEX(smc_server_lgr_pending);	/* serialize link group
 						 * creation on server

@@ -3273,9 +3274,17 @@ static int __init smc_init(void)
 		goto out_sock;
 	}
 
+	rc = smc_sysctl_init();
+	if (rc) {
+		pr_err("%s: sysctl_init fails with %d\n", __func__, rc);
+		goto out_ulp;
+	}
+
 	static_branch_enable(&tcp_have_smc);
 	return 0;
 
+out_ulp:
+	tcp_unregister_ulp(&smc_ulp_ops);
 out_sock:
 	sock_unregister(PF_SMC);
 out_proto6:

@@ -3303,6 +3312,7 @@ static int __init smc_init(void)
 static void __exit smc_exit(void)
 {
 	static_branch_disable(&tcp_have_smc);
+	smc_sysctl_exit();
 	tcp_unregister_ulp(&smc_ulp_ops);
 	sock_unregister(PF_SMC);
 	smc_core_exit();

diff --git a/net/smc/smc_sysctl.c b/net/smc/smc_sysctl.c
new file mode 100644
index 000000000000..8a3a8e145976
--- /dev/null
+++ b/net/smc/smc_sysctl.c

@@ -0,0 +1,70 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ *  Shared Memory Communications over RDMA (SMC-R) and RoCE
+ *
+ *  smc_sysctl.c: sysctl interface to SMC subsystem.
+ *
+ *  Copyright (c) 2022, Alibaba Inc.
+ *
+ *  Author: Tony Lu <tonylu@linux.alibaba.com>
+ *
+ */
+
+#include <linux/init.h>
+#include <linux/sysctl.h>
+#include <net/net_namespace.h>
+
+#include "smc_sysctl.h"
+
+static struct ctl_table smc_table[] = {
+	{  }
+};
+
+static __net_init int smc_sysctl_init_net(struct net *net)
+{
+	struct ctl_table *table;
+
+	table = smc_table;
+	if (!net_eq(net, &init_net)) {
+		int i;
+
+		table = kmemdup(table, sizeof(smc_table), GFP_KERNEL);
+		if (!table)
+			goto err_alloc;
+
+		for (i = 0; i < ARRAY_SIZE(smc_table) - 1; i++)
+			table[i].data += (void *)net - (void *)&init_net;
+	}
+
+	net->smc.smc_hdr = register_net_sysctl(net, "net/smc", table);
+	if (!net->smc.smc_hdr)
+		goto err_reg;
+
+	return 0;
+
+err_reg:
+	if (!net_eq(net, &init_net))
+		kfree(table);
+err_alloc:
+	return -ENOMEM;
+}
+
+static __net_exit void smc_sysctl_exit_net(struct net *net)
+{
+	unregister_net_sysctl_table(net->smc.smc_hdr);
+}
+
+static struct pernet_operations smc_sysctl_ops __net_initdata = {
+	.init = smc_sysctl_init_net,
+	.exit = smc_sysctl_exit_net,
+};
+
+int __init smc_sysctl_init(void)
+{
+	return register_pernet_subsys(&smc_sysctl_ops);
+}
+
+void smc_sysctl_exit(void)
+{
+	unregister_pernet_subsys(&smc_sysctl_ops);
+}

diff --git a/net/smc/smc_sysctl.h b/net/smc/smc_sysctl.h
new file mode 100644
index 000000000000..49553ac236b6
--- /dev/null
+++ b/net/smc/smc_sysctl.h

@@ -0,0 +1,32 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ *  Shared Memory Communications over RDMA (SMC-R) and RoCE
+ *
+ *  smc_sysctl.c: sysctl interface to SMC subsystem.
+ *
+ *  Copyright (c) 2022, Alibaba Inc.
+ *
+ *  Author: Tony Lu <tonylu@linux.alibaba.com>
+ *
+ */
+
+#ifndef _SMC_SYSCTL_H
+#define _SMC_SYSCTL_H
+
+#ifdef CONFIG_SYSCTL
+
+int smc_sysctl_init(void);
+void smc_sysctl_exit(void);
+
+#else
+
+int smc_sysctl_init(void)
+{
+	return 0;
+}
+
+void smc_sysctl_exit(void) { }
+
+#endif /* CONFIG_SYSCTL */
+
+#endif /* _SMC_SYSCTL_H */

-- 
2.19.1.3.ge56e4f7

This patch adds autocorking support for SMC which could improve
throughput for small message by x3+.

The main idea is borrowed from TCP autocorking with some RDMA
specific modification:
1. The first message should never cork to make sure we won't
   bring extra latency
2. If we have posted any Tx WRs to the NIC that have not
   completed, cork the new messages until:
   a) Receive CQE for the last Tx WR
   b) We have corked enough message on the connection
3. Try to push the corked data out when we receive CQE of
   the last Tx WR to prevent the corked messages hang in
   the send queue.

Both SMC autocorking and TCP autocorking check the TX completion
to decide whether we should cork or not. The difference is
when we got a SMC Tx WR completion, the data have been confirmed
by the RNIC while TCP TX completion just tells us the data
have been sent out by the local NIC.

Add an atomic variable tx_pushing in smc_connection to make
sure only one can send to let it cork more and save CDC slot.

SMC autocorking should not bring extra latency since the first
message will always been sent out immediately.

The qperf tcp_bw test shows more than x4 increase under small
message size with Mellanox connectX4-Lx, same result with other
throughput benchmarks like sockperf/netperf.
The qperf tcp_lat test shows SMC autocorking has not increase any
ping-pong latency.

Test command:
 client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \
			-t 30 -vu tcp_{bw|lat}
 server: smc_run taskset -c 1 qperf

=== Bandwidth ====
MsgSize(Bytes)  SMC-NoCork           TCP                      SMC-AutoCorking
      1         0.578 MB/s       2.392 MB/s(313.57%)        2.647 MB/s(357.72%)
      2         1.159 MB/s       4.780 MB/s(312.53%)        5.153 MB/s(344.71%)
      4         2.283 MB/s      10.266 MB/s(349.77%)       10.363 MB/s(354.02%)
      8         4.668 MB/s      19.040 MB/s(307.86%)       21.215 MB/s(354.45%)
     16         9.147 MB/s      38.904 MB/s(325.31%)       41.740 MB/s(356.32%)
     32        18.369 MB/s      79.587 MB/s(333.25%)       82.392 MB/s(348.52%)
     64        36.562 MB/s     148.668 MB/s(306.61%)      161.564 MB/s(341.89%)
    128        72.961 MB/s     274.913 MB/s(276.80%)      325.363 MB/s(345.94%)
    256       144.705 MB/s     512.059 MB/s(253.86%)      633.743 MB/s(337.96%)
    512       288.873 MB/s     884.977 MB/s(206.35%)     1250.681 MB/s(332.95%)
   1024       574.180 MB/s    1337.736 MB/s(132.98%)     2246.121 MB/s(291.19%)
   2048      1095.192 MB/s    1865.952 MB/s( 70.38%)     2057.767 MB/s( 87.89%)
   4096      2066.157 MB/s    2380.337 MB/s( 15.21%)     2173.983 MB/s(  5.22%)
   8192      3717.198 MB/s    2733.073 MB/s(-26.47%)     3491.223 MB/s( -6.08%)
  16384      4742.221 MB/s    2958.693 MB/s(-37.61%)     4637.692 MB/s( -2.20%)
  32768      5349.550 MB/s    3061.285 MB/s(-42.77%)     5385.796 MB/s(  0.68%)
  65536      5162.919 MB/s    3731.408 MB/s(-27.73%)     5223.890 MB/s(  1.18%)
==== Latency ====
MsgSize(Bytes)   SMC-NoCork         TCP                    SMC-AutoCorking
      1          10.540 us      11.938 us( 13.26%)       10.573 us(  0.31%)
      2          10.996 us      11.992 us(  9.06%)       10.269 us( -6.61%)
      4          10.229 us      11.687 us( 14.25%)       10.240 us(  0.11%)
      8          10.203 us      11.653 us( 14.21%)       10.402 us(  1.95%)
     16          10.530 us      11.313 us(  7.44%)       10.599 us(  0.66%)
     32          10.241 us      11.586 us( 13.13%)       10.223 us( -0.18%)
     64          10.693 us      11.652 us(  8.97%)       10.251 us( -4.13%)
    128          10.597 us      11.579 us(  9.27%)       10.494 us( -0.97%)
    256          10.409 us      11.957 us( 14.87%)       10.710 us(  2.89%)
    512          11.088 us      12.505 us( 12.78%)       10.547 us( -4.88%)
   1024          11.240 us      12.255 us(  9.03%)       10.787 us( -4.03%)
   2048          11.485 us      16.970 us( 47.76%)       11.256 us( -1.99%)
   4096          12.077 us      13.948 us( 15.49%)       12.230 us(  1.27%)
   8192          13.683 us      16.693 us( 22.00%)       13.786 us(  0.75%)
  16384          16.470 us      23.615 us( 43.38%)       16.459 us( -0.07%)
  32768          22.540 us      40.966 us( 81.75%)       23.284 us(  3.30%)
  65536          34.192 us      73.003 us(113.51%)       34.233 us(  0.12%)

With SMC autocorking support, we can archive better throughput
than TCP in most message sizes without any latency trade-off.

Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
---
v2: 1. Remove empty line in smc_connection
    2. use Reverse Christmas tree style for local variable.
    3. remove redundant container_of
v3: 1. use hex instead of decimal
    2. Remove unintented removal of new line
    3. Rename autocork to autocorking to be compliant with TCP
    4. re-test the data, use SMC NoCork as baseline
---
 net/smc/smc.h     |   2 +
 net/smc/smc_cdc.c |  11 +++--
 net/smc/smc_tx.c  | 107 ++++++++++++++++++++++++++++++++++++++++------
 3 files changed, 105 insertions(+), 15 deletions(-)

diff --git a/net/smc/smc.h b/net/smc/smc.h
index a096d8af21a0..e266b04b7585 100644
--- a/net/smc/smc.h
+++ b/net/smc/smc.h

@@ -29,6 +29,7 @@
 #define SMC_MAX_ISM_DEVS	8	/* max # of proposed non-native ISM
 					 * devices
 					 */
+#define SMC_AUTOCORKING_DEFAULT_SIZE	0x10000	/* 64K by default */
 
 extern struct proto smc_proto;
 extern struct proto smc_proto6;

@@ -192,6 +193,7 @@ struct smc_connection {
 						 * - dec on polled tx cqe
 						 */
 	wait_queue_head_t	cdc_pend_tx_wq; /* wakeup on no cdc_pend_tx_wr*/
+	atomic_t		tx_pushing;     /* nr_threads trying tx push */
 	struct delayed_work	tx_work;	/* retry of smc_cdc_msg_send */
 	u32			tx_off;		/* base offset in peer rmb */
 

diff --git a/net/smc/smc_cdc.c b/net/smc/smc_cdc.c
index 9d5a97168969..2b37bec90824 100644
--- a/net/smc/smc_cdc.c
+++ b/net/smc/smc_cdc.c

@@ -48,9 +48,14 @@ static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd,
 		conn->tx_cdc_seq_fin = cdcpend->ctrl_seq;
 	}
 
-	if (atomic_dec_and_test(&conn->cdc_pend_tx_wr) &&
-	    unlikely(wq_has_sleeper(&conn->cdc_pend_tx_wq)))
-		wake_up(&conn->cdc_pend_tx_wq);
+	if (atomic_dec_and_test(&conn->cdc_pend_tx_wr)) {
+		/* If this is the last pending WR complete, we must push to
+		 * prevent hang when autocork enabled.
+		 */
+		smc_tx_sndbuf_nonempty(conn);
+		if (unlikely(wq_has_sleeper(&conn->cdc_pend_tx_wq)))
+			wake_up(&conn->cdc_pend_tx_wq);
+	}
 	WARN_ON(atomic_read(&conn->cdc_pend_tx_wr) < 0);
 
 	smc_tx_sndbuf_nonfull(smc);

diff --git a/net/smc/smc_tx.c b/net/smc/smc_tx.c
index 436ac836f363..062c6b1535e3 100644
--- a/net/smc/smc_tx.c
+++ b/net/smc/smc_tx.c

@@ -131,6 +131,51 @@ static bool smc_tx_is_corked(struct smc_sock *smc)
 	return (tp->nonagle & TCP_NAGLE_CORK) ? true : false;
 }
 
+/* If we have pending CDC messages, do not send:
+ * Because CQE of this CDC message will happen shortly, it gives
+ * a chance to coalesce future sendmsg() payload in to one RDMA Write,
+ * without need for a timer, and with no latency trade off.
+ * Algorithm here:
+ *  1. First message should never cork
+ *  2. If we have pending Tx CDC messages, wait for the first CDC
+ *     message's completion
+ *  3. Don't cork to much data in a single RDMA Write to prevent burst
+ *     traffic, total corked message should not exceed sendbuf/2
+ */
+static bool smc_should_autocork(struct smc_sock *smc)
+{
+	struct smc_connection *conn = &smc->conn;
+	int corking_size;
+
+	corking_size = min(SMC_AUTOCORKING_DEFAULT_SIZE,
+			   conn->sndbuf_desc->len >> 1);
+
+	if (atomic_read(&conn->cdc_pend_tx_wr) == 0 ||
+	    smc_tx_prepared_sends(conn) > corking_size)
+		return false;
+	return true;
+}
+
+static bool smc_tx_should_cork(struct smc_sock *smc, struct msghdr *msg)
+{
+	struct smc_connection *conn = &smc->conn;
+
+	if (smc_should_autocork(smc))
+		return true;
+
+	/* for a corked socket defer the RDMA writes if
+	 * sndbuf_space is still available. The applications
+	 * should known how/when to uncork it.
+	 */
+	if ((msg->msg_flags & MSG_MORE ||
+	     smc_tx_is_corked(smc) ||
+	     msg->msg_flags & MSG_SENDPAGE_NOTLAST) &&
+	    atomic_read(&conn->sndbuf_space))
+		return true;
+
+	return false;
+}
+
 /* sndbuf producer: main API called by socket layer.
  * called under sock lock.
  */

@@ -235,13 +280,10 @@ int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len)
 		 */
 		if ((msg->msg_flags & MSG_OOB) && !send_remaining)
 			conn->urg_tx_pend = true;
-		/* for a corked socket defer the RDMA writes if
-		 * sndbuf_space is still available. The applications
-		 * should known how/when to uncork it.
+		/* If we need to cork, do nothing and wait for the next
+		 * sendmsg() call or push on tx completion
 		 */
-		if (!((msg->msg_flags & MSG_MORE || smc_tx_is_corked(smc) ||
-		       msg->msg_flags & MSG_SENDPAGE_NOTLAST) &&
-		      atomic_read(&conn->sndbuf_space)))
+		if (!smc_tx_should_cork(smc, msg))
 			smc_tx_sndbuf_nonempty(conn);
 
 		trace_smc_tx_sendmsg(smc, copylen);

@@ -589,13 +631,26 @@ static int smcd_tx_sndbuf_nonempty(struct smc_connection *conn)
 	return rc;
 }
 
-int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
+static int __smc_tx_sndbuf_nonempty(struct smc_connection *conn)
 {
-	int rc;
+	struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
+	int rc = 0;
+
+	/* No data in the send queue */
+	if (unlikely(smc_tx_prepared_sends(conn) <= 0))
+		goto out;
+
+	/* Peer don't have RMBE space */
+	if (unlikely(atomic_read(&conn->peer_rmbe_space) <= 0)) {
+		SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);
+		goto out;
+	}
 
 	if (conn->killed ||
-	    conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
-		return -EPIPE;	/* connection being aborted */
+	    conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
+		rc = -EPIPE;    /* connection being aborted */
+		goto out;
+	}
 	if (conn->lgr->is_smcd)
 		rc = smcd_tx_sndbuf_nonempty(conn);
 	else

@@ -603,10 +658,38 @@ int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
 
 	if (!rc) {
 		/* trigger socket release if connection is closing */
-		struct smc_sock *smc = container_of(conn, struct smc_sock,
-						    conn);
 		smc_close_wake_tx_prepared(smc);
 	}
+
+out:
+	return rc;
+}
+
+int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
+{
+	int rc;
+
+	/* This make sure only one can send simultaneously to prevent wasting
+	 * of CPU and CDC slot.
+	 * Record whether someone has tried to push while we are pushing.
+	 */
+	if (atomic_inc_return(&conn->tx_pushing) > 1)
+		return 0;
+
+again:
+	atomic_set(&conn->tx_pushing, 1);
+	smp_wmb(); /* Make sure tx_pushing is 1 before real send */
+	rc = __smc_tx_sndbuf_nonempty(conn);
+
+	/* We need to check whether someone else have added some data into
+	 * the send queue and tried to push but failed after the atomic_set()
+	 * when we are pushing.
+	 * If so, we need to push again to prevent those data hang in the send
+	 * queue.
+	 */
+	if (unlikely(!atomic_dec_and_test(&conn->tx_pushing)))
+		goto again;
+
 	return rc;
 }
 

-- 
2.19.1.3.ge56e4f7

This add a new sysctl: net.smc.autocorking_size

We can dynamically change the behaviour of autocorking
by change the value of autocorking_size.
Setting to 0 disables autocorking in SMC

Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
---
 Documentation/networking/smc-sysctl.rst | 23 +++++++++++++++++++++++
 include/net/netns/smc.h                 |  1 +
 net/smc/smc_sysctl.c                    | 10 ++++++++++
 net/smc/smc_tx.c                        |  2 +-
 4 files changed, 35 insertions(+), 1 deletion(-)
 create mode 100644 Documentation/networking/smc-sysctl.rst

diff --git a/Documentation/networking/smc-sysctl.rst b/Documentation/networking/smc-sysctl.rst
new file mode 100644
index 000000000000..c53f8c61c9e4
--- /dev/null
+++ b/Documentation/networking/smc-sysctl.rst

@@ -0,0 +1,23 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+=========
+SMC Sysctl
+=========
+
+/proc/sys/net/smc/* Variables
+==============================
+
+autocorking_size - INTEGER
+	Setting SMC auto corking size:
+	SMC auto corking is like TCP auto corking from the application's
+	perspective of view. When applications do consecutive small
+	write()/sendmsg() system calls, we try to coalesce these small writes
+	as much as possible, to lower total amount of CDC and RDMA Write been
+	sent.
+	autocorking_size limits the maximum corked bytes that can be sent to
+	the under device in 1 single sending. If set to 0, the SMC auto corking
+	is disabled.
+	Applications can still use TCP_CORK for optimal behavior when they
+	know how/when to uncork their sockets.
+
+	Default: 64K

diff --git a/include/net/netns/smc.h b/include/net/netns/smc.h
index 1682eae50579..e5389eeaf8bd 100644
--- a/include/net/netns/smc.h
+++ b/include/net/netns/smc.h

@@ -17,5 +17,6 @@ struct netns_smc {
 #ifdef CONFIG_SYSCTL
 	struct ctl_table_header		*smc_hdr;
 #endif
+	unsigned int			sysctl_autocorking_size;
 };
 #endif

diff --git a/net/smc/smc_sysctl.c b/net/smc/smc_sysctl.c
index 8a3a8e145976..3b59876aaac9 100644
--- a/net/smc/smc_sysctl.c
+++ b/net/smc/smc_sysctl.c

@@ -14,9 +14,17 @@
 #include <linux/sysctl.h>
 #include <net/net_namespace.h>
 
+#include "smc.h"
 #include "smc_sysctl.h"
 
 static struct ctl_table smc_table[] = {
+	{
+		.procname       = "autocorking_size",
+		.data           = &init_net.smc.sysctl_autocorking_size,
+		.maxlen         = sizeof(unsigned int),
+		.mode           = 0644,
+		.proc_handler	= proc_douintvec,
+	},
 	{  }
 };
 

@@ -40,6 +48,8 @@ static __net_init int smc_sysctl_init_net(struct net *net)
 	if (!net->smc.smc_hdr)
 		goto err_reg;
 
+	net->smc.sysctl_autocorking_size = SMC_AUTOCORKING_DEFAULT_SIZE;
+
 	return 0;
 
 err_reg:

diff --git a/net/smc/smc_tx.c b/net/smc/smc_tx.c
index 062c6b1535e3..257dc0d0aeb1 100644
--- a/net/smc/smc_tx.c
+++ b/net/smc/smc_tx.c

@@ -147,7 +147,7 @@ static bool smc_should_autocork(struct smc_sock *smc)
 	struct smc_connection *conn = &smc->conn;
 	int corking_size;
 
-	corking_size = min(SMC_AUTOCORKING_DEFAULT_SIZE,
+	corking_size = min(sock_net(&smc->sk)->smc.sysctl_autocorking_size,
 			   conn->sndbuf_desc->len >> 1);
 
 	if (atomic_read(&conn->cdc_pend_tx_wr) == 0 ||

-- 
2.19.1.3.ge56e4f7

rmbe_update_limit is used to limit announcing receive
window updating too frequently. RFC7609 request a minimal
increase in the window size of 10% of the receive buffer
space. But current implementation used:

  min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2)

and SOCK_MIN_SNDBUF / 2 == 2304 Bytes, which is almost
always less then 10% of the receive buffer space.

This causes the receiver always sending CDC message to
update its consumer cursor when it consumes more then 2K
of data. And as a result, we may encounter something like
"TCP silly window syndrome" when sending 2.5~8K message.

This patch fixes this using max(rmbe_size / 10, SOCK_MIN_SNDBUF / 2).

With this patch and SMC autocorking enabled, qperf 2K/4K/8K
tcp_bw test shows 45%/75%/40% increase in throughput respectively.

Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
---
 net/smc/smc_core.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/net/smc/smc_core.c b/net/smc/smc_core.c
index 29525d03b253..1674b2549f8b 100644
--- a/net/smc/smc_core.c
+++ b/net/smc/smc_core.c

@@ -1988,7 +1988,7 @@ static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
  */
 static inline int smc_rmb_wnd_update_limit(int rmbe_size)
 {
-	return min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
+	return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
 }
 
 /* map an rmb buf to a link */

-- 
2.19.1.3.ge56e4f7

When we are handling softirq workload, enable hardirq may
again interrupt the current routine of softirq, and then
try to raise softirq again. This only wastes CPU cycles
and won't have any real gain.

Since IB_CQ_REPORT_MISSED_EVENTS already make sure if
ib_req_notify_cq() returns 0, it is safe to wait for the
next event, with no need to poll the CQ again in this case.

This patch disables hardirq during the processing of softirq,
and re-arm the CQ after softirq is done. Somehow like NAPI.

Co-developed-by: Guangguan Wang <redacted>
Signed-off-by: Guangguan Wang <redacted>
Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
---
 net/smc/smc_wr.c | 49 +++++++++++++++++++++++++++---------------------
 1 file changed, 28 insertions(+), 21 deletions(-)

diff --git a/net/smc/smc_wr.c b/net/smc/smc_wr.c
index 24be1d03fef9..34d616406d51 100644
--- a/net/smc/smc_wr.c
+++ b/net/smc/smc_wr.c

@@ -137,25 +137,28 @@ static void smc_wr_tx_tasklet_fn(struct tasklet_struct *t)
 {
 	struct smc_ib_device *dev = from_tasklet(dev, t, send_tasklet);
 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
-	int i = 0, rc;
-	int polled = 0;
+	int i, rc;
 
 again:
-	polled++;
 	do {
 		memset(&wc, 0, sizeof(wc));
 		rc = ib_poll_cq(dev->roce_cq_send, SMC_WR_MAX_POLL_CQE, wc);
-		if (polled == 1) {
-			ib_req_notify_cq(dev->roce_cq_send,
-					 IB_CQ_NEXT_COMP |
-					 IB_CQ_REPORT_MISSED_EVENTS);
-		}
-		if (!rc)
-			break;
 		for (i = 0; i < rc; i++)
 			smc_wr_tx_process_cqe(&wc[i]);
+		if (rc < SMC_WR_MAX_POLL_CQE)
+			/* If < SMC_WR_MAX_POLL_CQE, the CQ should have been
+			 * drained, no need to poll again. --Guangguan Wang
+			 */
+			break;
 	} while (rc > 0);
-	if (polled == 1)
+
+	/* IB_CQ_REPORT_MISSED_EVENTS make sure if ib_req_notify_cq() returns
+	 * 0, it is safe to wait for the next event.
+	 * Else we must poll the CQ again to make sure we won't miss any event
+	 */
+	if (ib_req_notify_cq(dev->roce_cq_send,
+			     IB_CQ_NEXT_COMP |
+			     IB_CQ_REPORT_MISSED_EVENTS))
 		goto again;
 }
 

@@ -478,24 +481,28 @@ static void smc_wr_rx_tasklet_fn(struct tasklet_struct *t)
 {
 	struct smc_ib_device *dev = from_tasklet(dev, t, recv_tasklet);
 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
-	int polled = 0;
 	int rc;
 
 again:
-	polled++;
 	do {
 		memset(&wc, 0, sizeof(wc));
 		rc = ib_poll_cq(dev->roce_cq_recv, SMC_WR_MAX_POLL_CQE, wc);
-		if (polled == 1) {
-			ib_req_notify_cq(dev->roce_cq_recv,
-					 IB_CQ_SOLICITED_MASK
-					 | IB_CQ_REPORT_MISSED_EVENTS);
-		}
-		if (!rc)
+		if (rc > 0)
+			smc_wr_rx_process_cqes(&wc[0], rc);
+		if (rc < SMC_WR_MAX_POLL_CQE)
+			/* If < SMC_WR_MAX_POLL_CQE, the CQ should have been
+			 * drained, no need to poll again. --Guangguan Wang
+			 */
 			break;
-		smc_wr_rx_process_cqes(&wc[0], rc);
 	} while (rc > 0);
-	if (polled == 1)
+
+	/* IB_CQ_REPORT_MISSED_EVENTS make sure if ib_req_notify_cq() returns
+	 * 0, it is safe to wait for the next event.
+	 * Else we must poll the CQ again to make sure we won't miss any event
+	 */
+	if (ib_req_notify_cq(dev->roce_cq_recv,
+			     IB_CQ_SOLICITED_MASK |
+			     IB_CQ_REPORT_MISSED_EVENTS))
 		goto again;
 }
 

-- 
2.19.1.3.ge56e4f7

In commit ea785a1a573b("net/smc: Send directly when
TCP_CORK is cleared"), we don't use delayed work
to implement cork.

This patch use the same algorithm, removes the
delayed work when setting TCP_NODELAY and send
directly in setsockopt(). This also makes the
TCP_NODELAY the same as TCP.

Cc: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
---
 net/smc/af_smc.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
index 19b3066cf7af..e661b3747945 100644
--- a/net/smc/af_smc.c
+++ b/net/smc/af_smc.c

@@ -2796,8 +2796,8 @@ static int smc_setsockopt(struct socket *sock, int level, int optname,
 		    sk->sk_state != SMC_CLOSED) {
 			if (val) {
 				SMC_STAT_INC(smc, ndly_cnt);
-				mod_delayed_work(smc->conn.lgr->tx_wq,
-						 &smc->conn.tx_work, 0);
+				smc_tx_pending(&smc->conn);
+				cancel_delayed_work(&smc->conn.tx_work);
 			}
 		}
 		break;

-- 
2.19.1.3.ge56e4f7

Send data all the way down to the RDMA device is a time
consuming operation(get a new slot, maybe do RDMA Write
and send a CDC, etc). Moving those operations from BH
to user context is good for performance.

If the sock_lock is hold by user, we don't try to send
data out in the BH context, but just mark we should
send. Since the user will release the sock_lock soon, we
can do the sending there.

Add smc_release_cb() which will be called in release_sock()
and try send in the callback if needed.

This patch moves the sending part out from BH if sock lock
is hold by user. In my testing environment, this saves about
20% softirq in the qperf 4K tcp_bw test in the sender side
with no noticeable throughput drop.

Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
---
 net/smc/af_smc.c  | 16 ++++++++++++++++
 net/smc/smc.h     |  4 ++++
 net/smc/smc_cdc.c | 19 ++++++++++++++-----
 3 files changed, 34 insertions(+), 5 deletions(-)

diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
index e661b3747945..6447607675fa 100644
--- a/net/smc/af_smc.c
+++ b/net/smc/af_smc.c

@@ -193,12 +193,27 @@ void smc_unhash_sk(struct sock *sk)
 }
 EXPORT_SYMBOL_GPL(smc_unhash_sk);
 
+/* This will be called before user really release sock_lock. So do the
+ * work which we didn't do because of user hold the sock_lock in the
+ * BH context
+ */
+static void smc_release_cb(struct sock *sk)
+{
+	struct smc_sock *smc = smc_sk(sk);
+
+	if (smc->conn.tx_in_release_sock) {
+		smc_tx_pending(&smc->conn);
+		smc->conn.tx_in_release_sock = false;
+	}
+}
+
 struct proto smc_proto = {
 	.name		= "SMC",
 	.owner		= THIS_MODULE,
 	.keepalive	= smc_set_keepalive,
 	.hash		= smc_hash_sk,
 	.unhash		= smc_unhash_sk,
+	.release_cb	= smc_release_cb,
 	.obj_size	= sizeof(struct smc_sock),
 	.h.smc_hash	= &smc_v4_hashinfo,
 	.slab_flags	= SLAB_TYPESAFE_BY_RCU,

@@ -211,6 +226,7 @@ struct proto smc_proto6 = {
 	.keepalive	= smc_set_keepalive,
 	.hash		= smc_hash_sk,
 	.unhash		= smc_unhash_sk,
+	.release_cb	= smc_release_cb,
 	.obj_size	= sizeof(struct smc_sock),
 	.h.smc_hash	= &smc_v6_hashinfo,
 	.slab_flags	= SLAB_TYPESAFE_BY_RCU,

diff --git a/net/smc/smc.h b/net/smc/smc.h
index e266b04b7585..ea0620529ebe 100644
--- a/net/smc/smc.h
+++ b/net/smc/smc.h

@@ -213,6 +213,10 @@ struct smc_connection {
 						 * data still pending
 						 */
 	char			urg_rx_byte;	/* urgent byte */
+	bool			tx_in_release_sock;
+						/* flush pending tx data in
+						 * sock release_cb()
+						 */
 	atomic_t		bytes_to_rcv;	/* arrived data,
 						 * not yet received
 						 */

diff --git a/net/smc/smc_cdc.c b/net/smc/smc_cdc.c
index 2b37bec90824..5c731f27996e 100644
--- a/net/smc/smc_cdc.c
+++ b/net/smc/smc_cdc.c

@@ -49,10 +49,15 @@ static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd,
 	}
 
 	if (atomic_dec_and_test(&conn->cdc_pend_tx_wr)) {
-		/* If this is the last pending WR complete, we must push to
-		 * prevent hang when autocork enabled.
+		/* If user owns the sock_lock, mark the connection need sending.
+		 * User context will later try to send when it release sock_lock
+		 * in smc_release_cb()
 		 */
-		smc_tx_sndbuf_nonempty(conn);
+		if (sock_owned_by_user(&smc->sk))
+			conn->tx_in_release_sock = true;
+		else
+			smc_tx_pending(conn);
+
 		if (unlikely(wq_has_sleeper(&conn->cdc_pend_tx_wq)))
 			wake_up(&conn->cdc_pend_tx_wq);
 	}

@@ -355,8 +360,12 @@ static void smc_cdc_msg_recv_action(struct smc_sock *smc,
 	/* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
 	if ((diff_cons && smc_tx_prepared_sends(conn)) ||
 	    conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
-	    conn->local_rx_ctrl.prod_flags.urg_data_pending)
-		smc_tx_sndbuf_nonempty(conn);
+	    conn->local_rx_ctrl.prod_flags.urg_data_pending) {
+		if (!sock_owned_by_user(&smc->sk))
+			smc_tx_pending(conn);
+		else
+			conn->tx_in_release_sock = true;
+	}
 
 	if (diff_cons && conn->urg_tx_pend &&
 	    atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {

-- 
2.19.1.3.ge56e4f7

On Tue, Mar 01, 2022 at 05:44:01PM +0800, Dust Li wrote:

1. Please remove "--Guangguan Wang".
2. We already discussed that. SMC should be changed to use RDMA CQ pool API
drivers/infiniband/core/cq.c. 
ib_poll_handler() has much better implementation (tracing, IRQ rescheduling,
proper error handling) than this SMC variant.

Thanks

On Tue, Mar 01, 2022 at 12:14:15PM +0200, Leon Romanovsky wrote:

On Tue, Mar 01, 2022 at 05:44:01PM +0800, Dust Li wrote:

quoted

When we are handling softirq workload, enable hardirq may
again interrupt the current routine of softirq, and then
try to raise softirq again. This only wastes CPU cycles
and won't have any real gain.

Since IB_CQ_REPORT_MISSED_EVENTS already make sure if
ib_req_notify_cq() returns 0, it is safe to wait for the
next event, with no need to poll the CQ again in this case.

This patch disables hardirq during the processing of softirq,
and re-arm the CQ after softirq is done. Somehow like NAPI.

Co-developed-by: Guangguan Wang <redacted>
Signed-off-by: Guangguan Wang <redacted>
Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
---
 net/smc/smc_wr.c | 49 +++++++++++++++++++++++++++---------------------
 1 file changed, 28 insertions(+), 21 deletions(-)

diff --git a/net/smc/smc_wr.c b/net/smc/smc_wr.c
index 24be1d03fef9..34d616406d51 100644
--- a/net/smc/smc_wr.c
+++ b/net/smc/smc_wr.c

@@ -137,25 +137,28 @@ static void smc_wr_tx_tasklet_fn(struct tasklet_struct *t)
 {
 	struct smc_ib_device *dev = from_tasklet(dev, t, send_tasklet);
 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
-	int i = 0, rc;
-	int polled = 0;
+	int i, rc;
 
 again:
-	polled++;
 	do {
 		memset(&wc, 0, sizeof(wc));
 		rc = ib_poll_cq(dev->roce_cq_send, SMC_WR_MAX_POLL_CQE, wc);
-		if (polled == 1) {
-			ib_req_notify_cq(dev->roce_cq_send,
-					 IB_CQ_NEXT_COMP |
-					 IB_CQ_REPORT_MISSED_EVENTS);
-		}
-		if (!rc)
-			break;
 		for (i = 0; i < rc; i++)
 			smc_wr_tx_process_cqe(&wc[i]);
+		if (rc < SMC_WR_MAX_POLL_CQE)
+			/* If < SMC_WR_MAX_POLL_CQE, the CQ should have been
+			 * drained, no need to poll again. --Guangguan Wang

1. Please remove "--Guangguan Wang".
2. We already discussed that. SMC should be changed to use RDMA CQ pool API
drivers/infiniband/core/cq.c.
ib_poll_handler() has much better implementation (tracing, IRQ rescheduling,
proper error handling) than this SMC variant.

OK, I'll remove this patch in the next version.

Thanks

On Tue,  1 Mar 2022 17:43:58 +0800 Dust Li wrote:

I think this broke the build:

In file included from ../include/linux/kernel.h:26,
                 from ../include/linux/random.h:11,
                 from ../include/linux/net.h:18,
                 from ../net/smc/smc_tx.c:16:
../net/smc/smc_tx.c: In function ‘smc_should_autocork’:
../include/linux/minmax.h:20:35: error: comparison of distinct pointer types lacks a cast [-Werror]
   20 |         (!!(sizeof((typeof(x) *)1 == (typeof(y) *)1)))
      |                                   ^~
../include/linux/minmax.h:26:18: note: in expansion of macro ‘__typecheck’
   26 |                 (__typecheck(x, y) && __no_side_effects(x, y))
      |                  ^~~~~~~~~~~
../include/linux/minmax.h:36:31: note: in expansion of macro ‘__safe_cmp’
   36 |         __builtin_choose_expr(__safe_cmp(x, y), \
      |                               ^~~~~~~~~~
../include/linux/minmax.h:45:25: note: in expansion of macro ‘__careful_cmp’
   45 | #define min(x, y)       __careful_cmp(x, y, <)
      |                         ^~~~~~~~~~~~~
../net/smc/smc_tx.c:150:24: note: in expansion of macro ‘min’
  150 |         corking_size = min(sock_net(&smc->sk)->smc.sysctl_autocorking_size,
      |                        ^~~

On 01/03/2022 11:53, dust.li wrote:

On Tue, Mar 01, 2022 at 12:14:15PM +0200, Leon Romanovsky wrote:

quoted

On Tue, Mar 01, 2022 at 05:44:01PM +0800, Dust Li wrote:
1. Please remove "--Guangguan Wang".
2. We already discussed that. SMC should be changed to use RDMA CQ pool API
drivers/infiniband/core/cq.c.
ib_poll_handler() has much better implementation (tracing, IRQ rescheduling,
proper error handling) than this SMC variant.

OK, I'll remove this patch in the next version.

Looks like this one was accepted already, but per discussion (and I agree with that) -
please revert this patch. Thank you.

On Fri, Mar 04, 2022 at 09:19:27AM +0100, Karsten Graul wrote:

On 01/03/2022 11:53, dust.li wrote:

quoted

On Tue, Mar 01, 2022 at 12:14:15PM +0200, Leon Romanovsky wrote:

quoted

On Tue, Mar 01, 2022 at 05:44:01PM +0800, Dust Li wrote:
1. Please remove "--Guangguan Wang".
2. We already discussed that. SMC should be changed to use RDMA CQ pool API
drivers/infiniband/core/cq.c.
ib_poll_handler() has much better implementation (tracing, IRQ rescheduling,
proper error handling) than this SMC variant.

OK, I'll remove this patch in the next version.

Looks like this one was accepted already, but per discussion (and I agree with that) -
please revert this patch. Thank you.

Yes. No problem, I will send a revert patch.

Thanks