static inline void
send_burst(struct dpdk_knidev_writer *p)
{
uint32_t nb_tx;
nb_tx = rte_kni_tx_burst(p->kni, p->tx_buf, p->tx_buf_count);
DPDK_KNIDEV_WRITER_STATS_PKTS_DROP_ADD(p, p->tx_buf_count - nb_tx);
for ( ; nb_tx < p->tx_buf_count; nb_tx++)
/* TODO: a separate counter for this drop */
vr_dpdk_pfree(p->tx_buf[nb_tx], VP_DROP_INTERFACE_DROP);
p->tx_buf_count = 0;
}
static int
dpdk_knidev_writer_tx(void *port, struct rte_mbuf *pkt)
{
struct dpdk_knidev_writer *p = (struct dpdk_knidev_writer *) port;
struct rte_mbuf *pkt_copy;
/*
* KNI kernel module uses a trick to speed up packet processing. It takes
* a physical address of a memory pool, converts it to the kernel virtual
* address with phys_to_virt() and saves the address.
*
* Then in kni_net_rx_normal() instead of using phys_to_virt() per each
* packet, KNI just calculates the difference between the previously
* converted physical address of the given mempool and the packets
* physical address.
*
* It works well for the mbufs from the same mempool. It also works fine
* with any mempool allocated from the same physically contiguous memory
* segment.
*
* As soon as we get a mempool allocated from another memory segment, the
* difference calculations fail and thus we might have a crash.
*
* So we make sure the packet is from the RSS mempool. If not, we make
* a copy to the RSS mempool.
*/
if (unlikely(pkt->pool != vr_dpdk.rss_mempool ||
/* Check indirect mbuf's data is within the RSS mempool. */
rte_pktmbuf_mtod(pkt, uintptr_t) < vr_dpdk.rss_mempool->elt_va_start ||
rte_pktmbuf_mtod(pkt, uintptr_t) > vr_dpdk.rss_mempool->elt_va_end
)) {
pkt_copy = vr_dpdk_pktmbuf_copy(pkt, vr_dpdk.rss_mempool);
/* The original mbuf is no longer needed. */
vr_dpdk_pfree(pkt, VP_DROP_CLONED_ORIGINAL);
if (unlikely(pkt_copy == NULL)) {
DPDK_KNIDEV_WRITER_STATS_PKTS_DROP_ADD(p, 1);
return -1;
}
pkt = pkt_copy;
}
p->tx_buf[p->tx_buf_count++] = pkt;
DPDK_KNIDEV_WRITER_STATS_PKTS_IN_ADD(p, 1);
if (p->tx_buf_count >= p->tx_burst_sz)
send_burst(p);
return 0;
}
static void
vr_dpdk_packet_receive(struct vr_usocket *usockp)
{
const unsigned lcore_id = rte_lcore_id();
struct vr_dpdk_lcore *lcore = vr_dpdk.lcores[lcore_id];
struct vr_interface_stats *stats;
RTE_LOG(DEBUG, USOCK, "%s[%lx]: FD %d\n", __func__, pthread_self(),
usockp->usock_fd);
/**
* Packets is read from the agent's socket here. On success, a counter for
* packets dequeued from the interface is incremented.
*/
stats = vif_get_stats(usockp->usock_vif, lcore_id);
if (usockp->usock_vif) {
stats->vis_port_ipackets++;
/* buf_addr and data_off do not change */
usockp->usock_mbuf->data_len = usockp->usock_read_len;
usockp->usock_mbuf->pkt_len = usockp->usock_read_len;
/* convert mbuf to vr_packet */
vr_dpdk_packet_get(usockp->usock_mbuf, usockp->usock_vif);
/* send the mbuf to vRouter */
vr_dpdk_lcore_vroute(lcore, usockp->usock_vif, &usockp->usock_mbuf, 1);
/* flush packet TX queues immediately */
vr_dpdk_lcore_flush(lcore);
} else {
/**
* If reading from socket failed, increment counter for interface
* dequeue drops.
*/
RTE_LOG(ERR, VROUTER, "Error receiving from packet socket: no vif attached\n");
vr_dpdk_pfree(usockp->usock_mbuf, VP_DROP_INTERFACE_DROP);
stats->vis_port_ierrors++;
}
usockp->usock_mbuf = NULL;
usockp->usock_rx_buf = NULL;
usockp->usock_buf_len = 0;
return;
}
