ehea: NAPI multi queue TX/RX path for SMP
Subsystems (2)
PATCH 1/1 ehea: NAPI multi queue TX/RX path for SMP
Hi!
This patch replaces the former posted patch
([PATCH] ehea: Optional TX/RX path optimized for SMP) and provides a functionality that allows parallel RX processing on multiple RX queues by using dummy netdevices as proposed by Roland Dreier. The fairness amongst this net devices and other network devices is thus better than in our first approach.
Is there a way to avoid overloading a single CPU with multiple netdev receive processes using NAPI? A distribution based on CPU load might help to improve the throughput. Pinning interrupts shows there is still some room for improvements.
Regards,
Jan-Bernd
---
drivers/net/ehea/ehea.h +5 -6
@@ -39,7 +39,7 @@#include <asm/io.h> #define DRV_NAME "ehea" -#define DRV_VERSION "EHEA_0048" +#define DRV_VERSION "EHEA_0050"#define EHEA_MSG_DEFAULT (NETIF_MSG_LINK | NETIF_MSG_TIMER \ | NETIF_MSG_RX_ERR | NETIF_MSG_TX_ERR)
@@ -78,8 +78,6 @@#define EHEA_RQ2_PKT_SIZE 1522#define EHEA_L_PKT_SIZE 256/* low latency */ -#define EHEA_POLL_MAX_RWQE 1000 -/* Send completion signaling */ #define EHEA_SIG_IV_LONG 1
@@ -357,8 +355,8 @@ struct ehea_port_res {struct ehea_qp *qp;struct ehea_cq *send_cq;struct ehea_cq *recv_cq; -struct ehea_eq *send_eq; -struct ehea_eq *recv_eq; +struct ehea_eq *eq; +struct net_device *d_netdev;spinlock_t send_lock;struct ehea_q_skb_arr rq1_skba;struct ehea_q_skb_arr rq2_skba;
@@ -372,7 +370,6 @@ struct ehea_port_res {int swqe_count;u32 swqe_id_counter;u64 tx_packets; -struct tasklet_struct send_comp_task;spinlock_t recv_lock;struct port_state p_state;u64 rx_packets;
@@ -416,7 +413,9 @@ struct ehea_port {char int_aff_name[EHEA_IRQ_NAME_SIZE ];int allmulti;/* Indicates IFF_ALLMULTI state */ int promisc;/* Indicates IFF_PROMISC state */ +int num_tx_qps;int num_add_tx_qps; +int num_mcs;int resets;u64 mac_addr;u32 logical_port_id;
drivers/net/ehea/ehea_main.c +153 -123
536 lines not on this page · raw message
drivers/net/ehea/ehea_qmr.h +6 -1
@@ -320,6 +320,11 @@ static inline struct ehea_cqe *ehea_pollreturn hw_qeit_get_valid (queue); } +static inline void ehea_inc_cq (struct ehea_cq *cq) +{ +hw_qeit_inc (&cq->hw_queue); +} +static inline void ehea_inc_rq1 (struct ehea_qp *qp) {hw_qeit_inc (&qp->hw_rqueue1);
@@ -327,7 +332,7 @@ static inline void ehea_inc_rq1(struct estatic inline struct ehea_cqe *ehea_poll_cq (struct ehea_cq *my_cq) { -return hw_qeit_get_inc_valid (&my_cq->hw_queue); +return hw_qeit_get_valid (&my_cq->hw_queue); }#define EHEA_CQ_REGISTER_ORIG 0
Quoted 9 lines · Jan-Bernd Themann
Hi!
This patch replaces the former posted patch
([PATCH] ehea: Optional TX/RX path optimized for SMP) and provides a functionality that allows parallel RX processing on multiple RX queues by using dummy netdevices as proposed by Roland Dreier. The fairness amongst this net devices and other network devices is thus better than in our first approach.
Quoted 4 lines
Is there a way to avoid overloading a single CPU with multiple netdev receive processes using NAPI? A distribution based on CPU load might help to improve the throughput. Pinning interrupts shows there is still some room for improvements.
You may also want to look at the patches proposed by Dave Miller for splitting NAPI from netdevice so you can actually instanciate multiple NAPI "pollers"
Ben.