[1/2] net/bonding: fix a possible unbalance packet receiving
Checks
Commit Message
Current Rx round robin policy for the slaves has two issue:
1. active_slave in bond_dev_private is shared by multiple PMDS
which maybe cause some slave Rx hungry, for example, there
is two PMD and two slave port, both PMDs start to receive, and
see that active_slave is 0, and receive from slave 0, after
complete, they increase active_slave by one, totally active_slave
are increased by two, next time, they will start to receive
from slave 0 again, at last, slave 1 maybe drop packets during
to not be polled by PMD
2. active_slave is shared and written by multiple PMD in RX path
for every time RX, this is a kind of cache false share, low
performance.
so move active_slave from bond_dev_private to bond_rx_queue
make it as per queue variable
Signed-off-by: Li RongQing <lirongqing@baidu.com>
Signed-off-by: Dongsheng Rong <rongdongsheng@baidu.com>
---
drivers/net/bonding/eth_bond_private.h | 3 ++-
drivers/net/bonding/rte_eth_bond_api.c | 6 ------
drivers/net/bonding/rte_eth_bond_pmd.c | 14 +++++++-------
3 files changed, 9 insertions(+), 14 deletions(-)
Comments
Hi, Li RongQing
Please add fixes statements in the commit log.
And for the patch [2/2] of this series: Reviewed-by: Wei Hu (Xavier)
<xavier.huwei@huawei.com>
On 2020/9/30 13:40, Li RongQing wrote:
> ping
>
> thanks
>
> -Li
Fixes: ae2a04864a9a "(net/bonding: reduce slave starvation on Rx poll)"
Thanks
-Li
From: Wei Hu (Xavier) [mailto:huwei013@chinasoftinc.com]
Sent: Wednesday, September 30, 2020 4:41 PM
To: Li,Rongqing <lirongqing@baidu.com>
Cc: dev@dpdk.org; Wei Hu (Xavier) <xavier.huwei@huawei.com>
Subject: Re: [dpdk-dev] [PATCH] net/bonding: fix a possible unbalance packet receiving
Hi, Li RongQing
Please add fixes statements in the commit log.
And for the patch [2/2] of this series: Reviewed-by: Wei Hu (Xavier) <xavier.huwei@huawei.com><mailto:xavier.huwei@huawei.com>
Ping, should I resend this?
Thanks
-Li
On 10/9/2020 12:56 PM, Li,Rongqing wrote:
> Ping, should I resend this?
>
No, I can add Fixes line while merging, thanks.
On 9/22/2020 11:29 AM, Li RongQing wrote:
> Current Rx round robin policy for the slaves has two issue:
>
> 1. active_slave in bond_dev_private is shared by multiple PMDS
> which maybe cause some slave Rx hungry, for example, there
> is two PMD and two slave port, both PMDs start to receive, and
> see that active_slave is 0, and receive from slave 0, after
> complete, they increase active_slave by one, totally active_slave
> are increased by two, next time, they will start to receive
> from slave 0 again, at last, slave 1 maybe drop packets during
> to not be polled by PMD
>
> 2. active_slave is shared and written by multiple PMD in RX path
> for every time RX, this is a kind of cache false share, low
> performance.
>
> so move active_slave from bond_dev_private to bond_rx_queue
> make it as per queue variable
>
> Signed-off-by: Li RongQing <lirongqing@baidu.com>
> Signed-off-by: Dongsheng Rong <rongdongsheng@baidu.com>
>
Fixes: ae2a04864a9a ("net/bonding: reduce slave starvation on Rx poll")
Cc: stable@dpdk.org
For series,
Reviewed-by: Wei Hu (Xavier) <xavier.huwei@huawei.com>
Series applied to dpdk-next-net/main, thanks.
@@ -50,6 +50,8 @@ extern const struct rte_flow_ops bond_flow_ops;
/** Port Queue Mapping Structure */
struct bond_rx_queue {
uint16_t queue_id;
+ /**< Next active_slave to poll */
+ uint16_t active_slave;
/**< Queue Id */
struct bond_dev_private *dev_private;
/**< Reference to eth_dev private structure */
@@ -132,7 +134,6 @@ struct bond_dev_private {
uint16_t nb_rx_queues; /**< Total number of rx queues */
uint16_t nb_tx_queues; /**< Total number of tx queues*/
- uint16_t active_slave; /**< Next active_slave to poll */
uint16_t active_slave_count; /**< Number of active slaves */
uint16_t active_slaves[RTE_MAX_ETHPORTS]; /**< Active slave list */
@@ -129,12 +129,6 @@ deactivate_slave(struct rte_eth_dev *eth_dev, uint16_t port_id)
RTE_ASSERT(active_count < RTE_DIM(internals->active_slaves));
internals->active_slave_count = active_count;
- /* Resetting active_slave when reaches to max
- * no of slaves in active list
- */
- if (internals->active_slave >= active_count)
- internals->active_slave = 0;
-
if (eth_dev->data->dev_started) {
if (internals->mode == BONDING_MODE_8023AD) {
bond_mode_8023ad_start(eth_dev);
@@ -69,7 +69,7 @@ bond_ethdev_rx_burst(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
struct bond_rx_queue *bd_rx_q = (struct bond_rx_queue *)queue;
internals = bd_rx_q->dev_private;
slave_count = internals->active_slave_count;
- active_slave = internals->active_slave;
+ active_slave = bd_rx_q->active_slave;
for (i = 0; i < slave_count && nb_pkts; i++) {
uint16_t num_rx_slave;
@@ -86,8 +86,8 @@ bond_ethdev_rx_burst(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
active_slave = 0;
}
- if (++internals->active_slave >= slave_count)
- internals->active_slave = 0;
+ if (++bd_rx_q->active_slave >= slave_count)
+ bd_rx_q->active_slave = 0;
return num_rx_total;
}
@@ -303,9 +303,9 @@ rx_burst_8023ad(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts,
memcpy(slaves, internals->active_slaves,
sizeof(internals->active_slaves[0]) * slave_count);
- idx = internals->active_slave;
+ idx = bd_rx_q->active_slave;
if (idx >= slave_count) {
- internals->active_slave = 0;
+ bd_rx_q->active_slave = 0;
idx = 0;
}
for (i = 0; i < slave_count && num_rx_total < nb_pkts; i++) {
@@ -367,8 +367,8 @@ rx_burst_8023ad(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts,
idx = 0;
}
- if (++internals->active_slave >= slave_count)
- internals->active_slave = 0;
+ if (++bd_rx_q->active_slave >= slave_count)
+ bd_rx_q->active_slave = 0;
return num_rx_total;
}