No Honnappa. Thanks & Regards, Rajasekhar
On Wed, Mar 8, 2023 at 5:49 AM Honnappa Nagarahalli < honnappa.nagaraha...@arm.com> wrote: > > > From: Rajasekhar Pulluru <pullururajasek...@gmail.com> > Sent: Tuesday, March 7, 2023 12:52 PM > To: dev@dpdk.org > Subject: ixgbe rxq interrupt not working > > Hi Team, > > Bringing-up dpdk-22.07 on an intel machine with 8 ports, 4 of them driven > by igb and the rest of the 4 ports driven by ixgbe. > [Honnappa] Do you have packets crossing between the 2 drivers? > > > I am following the below sequence to initialize these ports: > > dev_conf.intr_conf.lsc = 1; //Enable link state change interrupt > dev_conf.intr_conf.rxq = 1; //Enable RX Queue Interrupt > dev_conf.rxmode.mq_mode = RTE_ETH_MQ_RX_NONE; > dev_conf.rxmode.offloads = 0; > dev_conf.txmode.mq_mode = RTE_ETH_MQ_TX_NONE; > dev_conf.txmode.offloads = 0; > > rte_eth_dev_configure > rte_eth_rx_queue_setup > rte_eth_tx_queue_setup > rte_eth_dev_start > data = port_id << CHAR_BIT | queue_id; > rte_eth_dev_rx_intr_ctl_q(port_id, queue_id, RTE_EPOLL_PER_THREAD, > RTE_INTR_EVENT_ADD, (void *)((uintptr_t)data)); > rte_eth_dev_rx_intr_enable(port_id, queue_id); > > And then main loop repeats the below: > > rte_epoll_wait(RTE_EPOLL_PER_THREAD, event, 1, timeout /* 200micro-sec > */); /* ignore return value */ > rte_eth_dev_rx_intr_disable(port_id, queue_id); > rte_eth_rx_burst(port_id, queue_id, pkts, num_pkts); > rte_eth_dev_rx_intr_enable(port_id, queue_id); > > The code is same for all the ports, igb ports are able to come-up and rx > packets, where-as the ixgbe ports are not able to rx packets at all. > cat /proc/interrupts dumps vfio-msix counters for ixgbe as 0, where-as > it's non-zero for igb. > If I don't use/enable rxq interrupt for ixgbe (and remove epoll wait, > interrupt enable/disable from while loop) and simply poll for > rte_eth_rx_burst in a loop, ixgbe ports are able to rx packets. > > What could be wrong here? Appreciate any help. > > I would also like to know if there's an asynchronous rxq interrupt > notification to the application instead of rte_epoll_wait (and sleep). > > Thanks & Regards, > Rajasekhar >
