[PATCH v3] net/bonding: reduce slave starvation on rx poll

[PATCH v3] net/bonding: reduce slave starvation on rx poll

[Keith Wiles]

When polling the bonded ports for RX packets the old driver would
always start with the first slave in the list. If the requested
number of packets is filled on the first port in a two port config
then the second port could be starved or have larger number of
missed packet errors.

The code attempts to start with a different slave each time RX poll
is done to help eliminate starvation of slave ports. The effect of
the previous code was much lower performance for two slaves in the
bond then just the one slave.

The performance drop was detected when the application can not poll
the rings of rx packets fast enough and the packets per second for
two or more ports was at the threshold thoughput of the application.
At this threshold the slaves would see very little or no drops in
the case of one slave. Then enable the second slave you would see
a large drop rate on the two slave bond and reduction in thoughput.

Signed-off-by: Keith Wiles <keith.wiles@intel.com>
---
v3 - remove more checkpatch errors
v2 - remove checkpatch errors

 drivers/net/bonding/rte_eth_bond_pmd.c     | 21 +++++++++++++++------
 drivers/net/bonding/rte_eth_bond_private.h |  3 ++-
 2 files changed, 17 insertions(+), 7 deletions(-)

diff --git a/drivers/net/bonding/rte_eth_bond_pmd.c b/drivers/net/bonding/rte_eth_bond_pmd.c
index f3ac9e273..18c7b2e85 100644
--- a/drivers/net/bonding/rte_eth_bond_pmd.c
+++ b/drivers/net/bonding/rte_eth_bond_pmd.c
@@ -1,7 +1,7 @@
 /*-
  *   BSD LICENSE
  *
- *   Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
+ *   Copyright(c) 2010-2017 Intel Corporation. All rights reserved.
  *   All rights reserved.
  *
  *   Redistribution and use in source and binary forms, with or without
@@ -145,7 +145,7 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
 	const uint16_t ether_type_slow_be = rte_be_to_cpu_16(ETHER_TYPE_SLOW);
 	uint16_t num_rx_total = 0;	/* Total number of received packets */
 	uint8_t slaves[RTE_MAX_ETHPORTS];
-	uint8_t slave_count;
+	uint8_t slave_count, idx;
 
 	uint8_t collecting;  /* current slave collecting status */
 	const uint8_t promisc = internals->promiscuous_en;
@@ -159,12 +159,18 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
 	memcpy(slaves, internals->active_slaves,
 			sizeof(internals->active_slaves[0]) * slave_count);
 
+	idx = internals->active_slave;
+	if (idx >= slave_count) {
+		internals->active_slave = 0;
+		idx = 0;
+	}
 	for (i = 0; i < slave_count && num_rx_total < nb_pkts; i++) {
 		j = num_rx_total;
-		collecting = ACTOR_STATE(&mode_8023ad_ports[slaves[i]], COLLECTING);
+		collecting = ACTOR_STATE(&mode_8023ad_ports[slaves[idx]],
+					 COLLECTING);
 
 		/* Read packets from this slave */
-		num_rx_total += rte_eth_rx_burst(slaves[i], bd_rx_q->queue_id,
+		num_rx_total += rte_eth_rx_burst(slaves[idx], bd_rx_q->queue_id,
 				&bufs[num_rx_total], nb_pkts - num_rx_total);
 
 		for (k = j; k < 2 && k < num_rx_total; k++)
@@ -187,8 +193,8 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
 					!is_same_ether_addr(&bond_mac, &hdr->d_addr)))) {
 
 				if (hdr->ether_type == ether_type_slow_be) {
-					bond_mode_8023ad_handle_slow_pkt(internals, slaves[i],
-						bufs[j]);
+					bond_mode_8023ad_handle_slow_pkt(
+					    internals, slaves[idx], bufs[j]);
 				} else
 					rte_pktmbuf_free(bufs[j]);
 
@@ -201,8 +207,11 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
 			} else
 				j++;
 		}
+		if (unlikely(++idx == slave_count))
+			idx = 0;
 	}
 
+	internals->active_slave = idx + 1;
 	return num_rx_total;
 }
 
diff --git a/drivers/net/bonding/rte_eth_bond_private.h b/drivers/net/bonding/rte_eth_bond_private.h
index 5a411e22b..c8db09005 100644
--- a/drivers/net/bonding/rte_eth_bond_private.h
+++ b/drivers/net/bonding/rte_eth_bond_private.h
@@ -1,7 +1,7 @@
 /*-
  *   BSD LICENSE
  *
- *   Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
+ *   Copyright(c) 2010-2017 Intel Corporation. All rights reserved.
  *   All rights reserved.
  *
  *   Redistribution and use in source and binary forms, with or without
@@ -144,6 +144,7 @@ struct bond_dev_private {
 	uint16_t nb_rx_queues;			/**< Total number of rx queues */
 	uint16_t nb_tx_queues;			/**< Total number of tx queues*/
 
+	uint8_t active_slave;		/**< Next active_slave to poll */
 	uint8_t active_slave_count;		/**< Number of active slaves */
 	uint8_t active_slaves[RTE_MAX_ETHPORTS];	/**< Active slave list */
 
-- 
2.12.0.rc2.2.g80ba04e

Re: [PATCH v3] net/bonding: reduce slave starvation on rx poll

[Declan Doherty]

Acking the correct version of the patch this time.

On 07/03/2017 10:39 PM, Keith Wiles wrote:
> When polling the bonded ports for RX packets the old driver would
> always start with the first slave in the list. If the requested
> number of packets is filled on the first port in a two port config
> then the second port could be starved or have larger number of
> missed packet errors.
>
> The code attempts to start with a different slave each time RX poll
> is done to help eliminate starvation of slave ports. The effect of
> the previous code was much lower performance for two slaves in the
> bond then just the one slave.
>
> The performance drop was detected when the application can not poll
> the rings of rx packets fast enough and the packets per second for
> two or more ports was at the threshold thoughput of the application.
> At this threshold the slaves would see very little or no drops in
> the case of one slave. Then enable the second slave you would see
> a large drop rate on the two slave bond and reduction in thoughput.
>
> Signed-off-by: Keith Wiles <keith.wiles@intel.com>
> ---
...
>
Acked-by: Declan Doherty <declan.doherty@intel.com>

Re: [PATCH v3] net/bonding: reduce slave starvation on rx poll

[Ferruh Yigit]

On 3/7/2017 10:39 PM, Keith Wiles wrote:
> When polling the bonded ports for RX packets the old driver would
> always start with the first slave in the list. If the requested
> number of packets is filled on the first port in a two port config
> then the second port could be starved or have larger number of
> missed packet errors.
> 
> The code attempts to start with a different slave each time RX poll
> is done to help eliminate starvation of slave ports. The effect of
> the previous code was much lower performance for two slaves in the
> bond then just the one slave.
> 
> The performance drop was detected when the application can not poll
> the rings of rx packets fast enough and the packets per second for
> two or more ports was at the threshold thoughput of the application.
> At this threshold the slaves would see very little or no drops in
> the case of one slave. Then enable the second slave you would see
> a large drop rate on the two slave bond and reduction in thoughput.
> 
> Signed-off-by: Keith Wiles <keith.wiles@intel.com>
> ---
> v3 - remove more checkpatch errors
> v2 - remove checkpatch errors
> 
>  drivers/net/bonding/rte_eth_bond_pmd.c     | 21 +++++++++++++++------
>  drivers/net/bonding/rte_eth_bond_private.h |  3 ++-
>  2 files changed, 17 insertions(+), 7 deletions(-)
> 
> diff --git a/drivers/net/bonding/rte_eth_bond_pmd.c b/drivers/net/bonding/rte_eth_bond_pmd.c
> index f3ac9e273..18c7b2e85 100644
> --- a/drivers/net/bonding/rte_eth_bond_pmd.c
> +++ b/drivers/net/bonding/rte_eth_bond_pmd.c
> @@ -1,7 +1,7 @@
>  /*-
>   *   BSD LICENSE
>   *
> - *   Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
> + *   Copyright(c) 2010-2017 Intel Corporation. All rights reserved.
>   *   All rights reserved.
>   *
>   *   Redistribution and use in source and binary forms, with or without
> @@ -145,7 +145,7 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
>  	const uint16_t ether_type_slow_be = rte_be_to_cpu_16(ETHER_TYPE_SLOW);
>  	uint16_t num_rx_total = 0;	/* Total number of received packets */
>  	uint8_t slaves[RTE_MAX_ETHPORTS];
> -	uint8_t slave_count;
> +	uint8_t slave_count, idx;
>  
>  	uint8_t collecting;  /* current slave collecting status */
>  	const uint8_t promisc = internals->promiscuous_en;
> @@ -159,12 +159,18 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
>  	memcpy(slaves, internals->active_slaves,
>  			sizeof(internals->active_slaves[0]) * slave_count);
>  
> +	idx = internals->active_slave;
> +	if (idx >= slave_count) {
> +		internals->active_slave = 0;
> +		idx = 0;
> +	}
>  	for (i = 0; i < slave_count && num_rx_total < nb_pkts; i++) {
>  		j = num_rx_total;
> -		collecting = ACTOR_STATE(&mode_8023ad_ports[slaves[i]], COLLECTING);
> +		collecting = ACTOR_STATE(&mode_8023ad_ports[slaves[idx]],
> +					 COLLECTING);
>  
>  		/* Read packets from this slave */
> -		num_rx_total += rte_eth_rx_burst(slaves[i], bd_rx_q->queue_id,
> +		num_rx_total += rte_eth_rx_burst(slaves[idx], bd_rx_q->queue_id,
>  				&bufs[num_rx_total], nb_pkts - num_rx_total);
>  
>  		for (k = j; k < 2 && k < num_rx_total; k++)
> @@ -187,8 +193,8 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
>  					!is_same_ether_addr(&bond_mac, &hdr->d_addr)))) {
>  
>  				if (hdr->ether_type == ether_type_slow_be) {
> -					bond_mode_8023ad_handle_slow_pkt(internals, slaves[i],
> -						bufs[j]);
> +					bond_mode_8023ad_handle_slow_pkt(
> +					    internals, slaves[idx], bufs[j]);
>  				} else
>  					rte_pktmbuf_free(bufs[j]);
>  
> @@ -201,8 +207,11 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
>  			} else
>  				j++;
>  		}
> +		if (unlikely(++idx == slave_count))
> +			idx = 0;
>  	}
>  
> +	internals->active_slave = idx + 1;

idx already increased above, won't using "idx + 1" here cause skipping
some slaves in some conditions depending num_rx_total?

Like,

slave_count = 2
internals->active_slave = 0;

idx = 0; receive from slave[0]
idx = 1;
num_rx_total >= nb_pkts
internals->active_slave = 2;

idx = 2;
internals->active_slave = 0;
idx = 0; receive from slave[0]
idx = 1;
num_rx_total >= nb_pkts
internals->active_slave = 2;

....

Always read from slave[0] as long as it fills num_rx_total threshold,
same as previous.

OR

slave_count = 3
internals->active_slave = 0;

idx = 0; receive from slave[0]
idx = 1; receive from slave[1]
idx = 2;
num_rx_total >= nb_pkts
internals->active_slave = 3;

idx = 3;
internals->active_slave = 0;
idx = 0; receive from slave[0]
idx = 1; receive from slave[1]
idx = 2;
num_rx_total >= nb_pkts
internals->active_slave = 3;

....

Not read from slave[2] as long as slave[0 & 1] fills num_rx_total threshold.

>  	return num_rx_total;
>  }
>  
> diff --git a/drivers/net/bonding/rte_eth_bond_private.h b/drivers/net/bonding/rte_eth_bond_private.h
> index 5a411e22b..c8db09005 100644
> --- a/drivers/net/bonding/rte_eth_bond_private.h
> +++ b/drivers/net/bonding/rte_eth_bond_private.h
> @@ -1,7 +1,7 @@
>  /*-
>   *   BSD LICENSE
>   *
> - *   Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
> + *   Copyright(c) 2010-2017 Intel Corporation. All rights reserved.
>   *   All rights reserved.
>   *
>   *   Redistribution and use in source and binary forms, with or without
> @@ -144,6 +144,7 @@ struct bond_dev_private {
>  	uint16_t nb_rx_queues;			/**< Total number of rx queues */
>  	uint16_t nb_tx_queues;			/**< Total number of tx queues*/
>  
> +	uint8_t active_slave;		/**< Next active_slave to poll */
>  	uint8_t active_slave_count;		/**< Number of active slaves */
>  	uint8_t active_slaves[RTE_MAX_ETHPORTS];	/**< Active slave list */
>  
> 

[PATCH v4] net/bonding: reduce slave starvation on rx poll

[Keith Wiles]

When polling the bonded ports for RX packets the old driver would
always start with the first slave in the list. If the requested
number of packets is filled on the first port in a two port config
then the second port could be starved or have larger number of
missed packet errors.

The code attempts to start with a different slave each time RX poll
is done to help eliminate starvation of slave ports. The effect of
the previous code was much lower performance for two slaves in the
bond then just the one slave.

The performance drop was detected when the application can not poll
the rings of rx packets fast enough and the packets per second for
two or more ports was at the threshold thoughput of the application.
At this threshold the slaves would see very little or no drops in
the case of one slave. Then enable the second slave you would see
a large drop rate on the two slave bond and reduction in thoughput.

Signed-off-by: Keith Wiles <keith.wiles@intel.com>
---
v4 - fix saving of next slave to reduce skipping a slave.
v3 - remove more checkpatch errors
v2 - remove checkpatch errors

 drivers/net/bonding/rte_eth_bond_pmd.c     | 21 +++++++++++++++------
 drivers/net/bonding/rte_eth_bond_private.h |  3 ++-
 2 files changed, 17 insertions(+), 7 deletions(-)

diff --git a/drivers/net/bonding/rte_eth_bond_pmd.c b/drivers/net/bonding/rte_eth_bond_pmd.c
index f3ac9e273..bc6e277ef 100644
--- a/drivers/net/bonding/rte_eth_bond_pmd.c
+++ b/drivers/net/bonding/rte_eth_bond_pmd.c
@@ -1,7 +1,7 @@
 /*-
  *   BSD LICENSE
  *
- *   Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
+ *   Copyright(c) 2010-2017 Intel Corporation. All rights reserved.
  *   All rights reserved.
  *
  *   Redistribution and use in source and binary forms, with or without
@@ -145,7 +145,7 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
 	const uint16_t ether_type_slow_be = rte_be_to_cpu_16(ETHER_TYPE_SLOW);
 	uint16_t num_rx_total = 0;	/* Total number of received packets */
 	uint8_t slaves[RTE_MAX_ETHPORTS];
-	uint8_t slave_count;
+	uint8_t slave_count, idx;
 
 	uint8_t collecting;  /* current slave collecting status */
 	const uint8_t promisc = internals->promiscuous_en;
@@ -159,12 +159,18 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
 	memcpy(slaves, internals->active_slaves,
 			sizeof(internals->active_slaves[0]) * slave_count);
 
+	idx = internals->active_slave;
+	if (idx >= slave_count) {
+		internals->active_slave = 0;
+		idx = 0;
+	}
 	for (i = 0; i < slave_count && num_rx_total < nb_pkts; i++) {
 		j = num_rx_total;
-		collecting = ACTOR_STATE(&mode_8023ad_ports[slaves[i]], COLLECTING);
+		collecting = ACTOR_STATE(&mode_8023ad_ports[slaves[idx]],
+					 COLLECTING);
 
 		/* Read packets from this slave */
-		num_rx_total += rte_eth_rx_burst(slaves[i], bd_rx_q->queue_id,
+		num_rx_total += rte_eth_rx_burst(slaves[idx], bd_rx_q->queue_id,
 				&bufs[num_rx_total], nb_pkts - num_rx_total);
 
 		for (k = j; k < 2 && k < num_rx_total; k++)
@@ -187,8 +193,8 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
 					!is_same_ether_addr(&bond_mac, &hdr->d_addr)))) {
 
 				if (hdr->ether_type == ether_type_slow_be) {
-					bond_mode_8023ad_handle_slow_pkt(internals, slaves[i],
-						bufs[j]);
+					bond_mode_8023ad_handle_slow_pkt(
+					    internals, slaves[idx], bufs[j]);
 				} else
 					rte_pktmbuf_free(bufs[j]);
 
@@ -201,8 +207,11 @@ bond_ethdev_rx_burst_8023ad(void *queue, struct rte_mbuf **bufs,
 			} else
 				j++;
 		}
+		if (unlikely(++idx == slave_count))
+			idx = 0;
 	}
 
+	internals->active_slave = idx;
 	return num_rx_total;
 }
 
diff --git a/drivers/net/bonding/rte_eth_bond_private.h b/drivers/net/bonding/rte_eth_bond_private.h
index 5a411e22b..c8db09005 100644
--- a/drivers/net/bonding/rte_eth_bond_private.h
+++ b/drivers/net/bonding/rte_eth_bond_private.h
@@ -1,7 +1,7 @@
 /*-
  *   BSD LICENSE
  *
- *   Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
+ *   Copyright(c) 2010-2017 Intel Corporation. All rights reserved.
  *   All rights reserved.
  *
  *   Redistribution and use in source and binary forms, with or without
@@ -144,6 +144,7 @@ struct bond_dev_private {
 	uint16_t nb_rx_queues;			/**< Total number of rx queues */
 	uint16_t nb_tx_queues;			/**< Total number of tx queues*/
 
+	uint8_t active_slave;		/**< Next active_slave to poll */
 	uint8_t active_slave_count;		/**< Number of active slaves */
 	uint8_t active_slaves[RTE_MAX_ETHPORTS];	/**< Active slave list */
 
-- 
2.12.0.rc2.2.g80ba04e

Re: [PATCH v4] net/bonding: reduce slave starvation on rx poll

[Ferruh Yigit]

On 3/21/2017 3:12 PM, Keith Wiles wrote:
> When polling the bonded ports for RX packets the old driver would
> always start with the first slave in the list. If the requested
> number of packets is filled on the first port in a two port config
> then the second port could be starved or have larger number of
> missed packet errors.
> 
> The code attempts to start with a different slave each time RX poll
> is done to help eliminate starvation of slave ports. The effect of
> the previous code was much lower performance for two slaves in the
> bond then just the one slave.
> 
> The performance drop was detected when the application can not poll
> the rings of rx packets fast enough and the packets per second for
> two or more ports was at the threshold thoughput of the application.
> At this threshold the slaves would see very little or no drops in
> the case of one slave. Then enable the second slave you would see
> a large drop rate on the two slave bond and reduction in thoughput.
> 
> Signed-off-by: Keith Wiles <keith.wiles@intel.com>

Acked-by: Declan Doherty <declan.doherty@intel.com>

Applied to dpdk-next-net/master, thanks.