net/rds: Fix endless RNR situation

Working with the following SRs:

	Exadata SR# 3-15640329311
	Linux SR#3-15675579325

it was discovered that by inserting IB_SEND_SOLICITED at regular
intervals removed the endless RNR Retry situation. The test was made
by inserting IB_SEND_SOLICITED at the same interval as
IB_SEND_SIGNALED was inserted, that is, by default for every 17th
fragment.

This commit introduces the sysctl variable
net.rds.ib.max_unsolicited_wr. A value of zero disables the
functionality of inserting IB_SEND_SOLICITED. A value of N will insert
IB_SEND_SOLICITED for every Nth fragment.

net.rds.ib.max_unsolicited_wr is by default 16, in order to avoid
customization when this fix is applied at the customer site.

This fix also has the nice side-effect that it improves IOPS for 1Q,
1D, 1T cases:

-q 1M -a 256:

Without fix:

tsks   tx/s   rx/s  tx+rx K/s    mbi K/s    mbo K/s tx us/c   rtt us cpu %
   1   1161      0 1189243.20       0.00       0.00  203.52   857.34 -1.00
(average)

With fix (with default net.rds.ib.max_unsolicited_wr = 16):

tsks   tx/s   rx/s  tx+rx K/s    mbi K/s    mbo K/s tx us/c   rtt us cpu %
   1   1323      0 1355849.36       0.00       0.00  203.76   751.50 -1.00
(average)

-q $[32*1024+256] -a 256:

With fix (net.rds.ib.max_unsolicited_wr = 0, i.e. disabled):

tsks   tx/s   rx/s  tx+rx K/s    mbi K/s    mbo K/s tx us/c   rtt us cpu %
   1  15243      0  492547.75       0.00       0.00   10.58    62.01 -1.00
(average)

Ditto with net.rds.ib.max_unsolicited_wr = 4 (two SEND_SOLICITED per ~32K):

tsks   tx/s   rx/s  tx+rx K/s    mbi K/s    mbo K/s tx us/c   rtt us cpu %
   1  16422      0  530641.03       0.00       0.00   10.28    57.25 -1.00
(average)

Orabug: 28856912

Reviewed-by: Håkon Bugge <[email protected]>
Signed-off-by: Venkat Venkatsubra <[email protected]>

Author: venkatxvenkatsubra

net/rds/ib.h

@@ -259,6 +259,9 @@ struct rds_ib_connection {
 
 	/* Batched completions */
 	unsigned int		i_unsignaled_wrs;
+
+	/* Wake up receiver once in a while */
+	unsigned int		i_unsolicited_wrs;
 	u8                      i_sl;
 
 	atomic_t                i_cache_allocs;
@@ -600,6 +603,7 @@ void rds_ib_sysctl_exit(void);
 extern unsigned long rds_ib_sysctl_max_send_wr;
 extern unsigned long rds_ib_sysctl_max_recv_wr;
 extern unsigned long rds_ib_sysctl_max_unsig_wrs;
+extern unsigned long rds_ib_sysctl_max_unsolicited_wrs;
 extern unsigned long rds_ib_sysctl_max_unsig_bytes;
 extern unsigned long rds_ib_sysctl_max_recv_allocation;
 extern unsigned int rds_ib_sysctl_flow_control;

net/rds/ib_send.c

@@ -538,9 +538,15 @@ static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
 	if (ic->i_unsignaled_wrs-- == 0 || notify) {
 		ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 		send->s_wr.send_flags |= IB_SEND_SIGNALED;
-		return 1;
 	}
-	return 0;
+
+	/* To keep the rx pipeline going, add SEND_SOLIICITED once in a while */
+	if (rds_ib_sysctl_max_unsolicited_wrs && --ic->i_unsolicited_wrs == 0) {
+		ic->i_unsolicited_wrs = rds_ib_sysctl_max_unsolicited_wrs;
+		send->s_wr.send_flags |= IB_SEND_SOLICITED;
+	}
+
+	return !!(send->s_wr.send_flags & IB_SEND_SIGNALED);
 }
 
 /*
@@ -645,6 +651,7 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 			rm->data.op_count = 0;
 		}
 
+		ic->i_unsolicited_wrs = rds_ib_sysctl_max_unsolicited_wrs;
 		rds_message_addref(rm);
 		rm->data.op_dmasg = 0;
 		rm->data.op_dmaoff = 0;

net/rds/ib_sysctl.c

@@ -49,6 +49,16 @@ unsigned long rds_ib_sysctl_max_unsig_wrs = 16;
 static unsigned long rds_ib_sysctl_max_unsig_wr_min = 1;
 static unsigned long rds_ib_sysctl_max_unsig_wr_max = 64;
 
+unsigned long rds_ib_sysctl_max_unsolicited_wrs = 16;
+
+/* Zero means inserting SEND_SOLICITED in the middle of an RDS message
+ * is disabled
+ */
+static unsigned long rds_ib_sysctl_max_unsolicited_wr_min;
+/* Nmbr frags of 1MB + 256B RDBMS hdr */
+static unsigned long rds_ib_sysctl_max_unsolicited_wr_max =
+	(1 * 1024 * 1024 + RDS_FRAG_SIZE) / RDS_FRAG_SIZE;
+
 /*
  * This sysctl does nothing.
  *
@@ -91,6 +101,15 @@ static struct ctl_table rds_ib_sysctl_table[] = {
 		.extra1		= &rds_ib_sysctl_max_unsig_wr_min,
 		.extra2		= &rds_ib_sysctl_max_unsig_wr_max,
 	},
+	{
+		.procname       = "max_unsolicited_wr",
+		.data		= &rds_ib_sysctl_max_unsolicited_wrs,
+		.maxlen         = sizeof(unsigned long),
+		.mode           = 0644,
+		.proc_handler   = &proc_doulongvec_minmax,
+		.extra1		= &rds_ib_sysctl_max_unsolicited_wr_min,
+		.extra2		= &rds_ib_sysctl_max_unsolicited_wr_max,
+	},
 	{
 		.procname       = "max_recv_allocation",
 		.data		= &rds_ib_sysctl_max_recv_allocation,