dpdk-kcp

A kcp wrapper DPDK library.

How to use

1. Put the source code to lib and modify the meson.build.

cp librte_ikcp $RTE_SDK/lib
# Add 'ikcp' to libraries.
vi $RTE_SDK/lib/meson.build

2. Compile the lib.

cd $RTE_SDK/build
ninja

API

1. rte_ikcp data structure. A rte_ikcp structure contains internal ikcp structure, rte_ikcp_config structure and rx/tx queues used for multi-queue mode.

/**
 * The RTE ikcp structure.
 */
struct rte_ikcp {
	ikcpcb kcp;                           /**< Internal kcp object. */
	struct rte_ikcp_config config;        /**< Internal config. */
	struct rte_ring *rx_queue, *tx_queue; /**< RX/TX queues when RTE_IKCP_MQ is enabled. */
};

2. rte_ikcp_config data structure. This structure mainly describe what addresses, which port and which queue would be used to send. Flags are defined in rte_ikcp.h, for example flags with RTE_IKCP_MQ will enable multi-queue mode.

/**
 * The RTE ikcp config structure.
 */
struct rte_ikcp_config {
	uint16_t port_id;                   /**< Port id. */
	uint16_t txq_id;                    /**< TX queue id of the port, used when RTE_IKCP_MQ is disabled. */
	union {
		uint32_t ipv4;
		uint8_t ipv6[16];
	} src_ip, dst_ip;                   /**< Addresses to send and receive. */
	uint16_t src_port;                  /**< UDP port when sending. */
	uint16_t dst_port;                  /**< UDP port when receiving. */
	struct rte_ether_addr self_mac;     /**< MAC address of the port. */
	struct rte_ether_addr next_hop_mac; /**< MAC address of the next hop. */
	struct rte_mempool *mempool;        /**< The mempool to allocate mbufs. */
	uint32_t flags;                     /**< Config flags. */
};

3. Create a rte_ikcp object. This API create a rte_ikcp object with a given rte_ikcp_config, the conv and the user is used in internal ikcp_create, see ikcp.h and ikcp.c for more information. A rte_ickp is associated with a certain port after creating.

struct rte_ikcp *rte_ikcp_create(struct rte_ikcp_config config, uint32_t conv, void *user);

4. Release a rte_ikcp object. This API will release a rte_ikcp object which is created before. When a port is released the rte_ikcp will be release as well by the RTE_ETH_EVENT_DESTROY callback.

void rte_ikcp_release(struct rte_ikcp *ikcp);

5. Update internal kcp status. This API will pass current timestamp to kcp_update. It should be invoked around every 10~100ms. Or use rte_ikcp_check to decied when to invoke.

void rte_ikcp_update(struct rte_ikcp *ikcp);

6. Get the next tsc cycle when rte_ikcp_update should be invoked.

uint64_t rte_ikcp_check(struct rte_ikcp *ikcp);

7. Send data to kcp. This API will send data to internal kcp tx queue. The data will be sent by rte_ikcp_flush or rte_ikcp_update on most occations.

int32_t rte_ikcp_send(struct rte_ikcp *ikcp, const char* data, int32_t len);

8. Recv data from kcp. This API will recv data from kcp.

int32_t rte_ikcp_recv(struct rte_ikcp *ikcp, char* data, int32_t len);

9. Input data to internal kcp rx queue in single-queue mode. The mbuf is checked whether belongs to the rte_ikcp or not and freed after enqueue sucessfully.

int32_t rte_ikcp_input(struct rte_ikcp *ikcp, struct rte_mbuf *mbuf);

10. Input data from multi mbufs from rte_ikcp->rx_queue in multi-queue mode. The mbufs should be enqueued to the rte_ikcp->rx_queue already.

int32_t rte_ikcp_input_bulk(struct rte_ikcp *ikcp);

11. Flush pending data of kcp. This API directly invoke ikcp_flush. It should not be invoked directly on most occations.

void rte_ikcp_flush(struct rte_ikcp *ikcp);

12. Set the internal params of a kcp. This API directly invoke ikcp_nodelay. See ikcp.h for more information.

void rte_ikcp_nodelay(struct rte_ikcp *ikcp, int32_t nodelay, int32_t interval, int32_t resend, int32_t nc);

13. Check the size of next message in the internal kcp rx queue. This API directly invoke ikcp_peeksize. See ikcp.h for more information.

int32_t rte_ikcp_peeksize(struct rte_ikcp *ikcp);

14. Set the internal MTU of a kcp. This API directly invoke ikcp_setmtu and will be invoked in rte_ikcp_create.

int32_t rte_ikcp_setmtu(struct rte_ikcp *ikcp, int32_t mtu);

15. Set the internal windows size of a kcp. This API directly invoke ikcp_wndsize. See ikcp.h for more information.

int32_t rte_ikcp_wndsize(struct rte_ikcp *ikcp, int32_t sndwnd, int32_t rcvwnd);

16. Get how many packet is waiting to be sent of a kcp. This API directly invoke ikcp_waitsnd. See ikcp.h for more information.

int32_t rte_ikcp_waitsnd(struct rte_ikcp *ikcp);

Example

The example contains 2 modes with different kcp backend. Mode can be changed with --dpdk-kcp or --socket-kcp. If you use linux kcp backend, the packet send/recv will use socket rather than librte_ikcp.

[root@localhost examples]# ./dpdk-ikcp -h
Usage: ./dpdk-ikcp [options]
options:
  --socket-kcp: Use kcp socket backend
  --dpdk-kcp: Use kcp pmd backend

The available arguments are here:

# dpdk backend.
[root@localhost examples]# ./dpdk-ikcp --dpdk-kcp -- -h
./dpdk-ikcp --dpdk-kcp [EAL options] -- [--debug] [-4|-6] --device PCI --src SRC --dst DST --next-hop-mac M:M:M:M:M:M
                -4: Use IPv4 (default)
                -6: Use IPv6
                -v: Display verbose send/recv details
                --device: Assign the PCI address of device
                --src: Assign the source IP
                --dst: Assign the destination IP
                --next-hop-mac: Assign the MAC address of the next hop
                --multi-queue: Use multi queues to transmit
                --debug: Enable kcp private log
                
# socket backend.
[root@localhost examples]# ./dpdk-ikcp --socket-kcp -h
./dpdk-ikcp --socket-kcp [-4|-6] --dst DST
                -4: IP address is IPv4 (default)
                -6: IP address is IPv6
                -v: Display verbose send/recv details
                --src: Assign the source IP
                --dst: Assign the destination IP

And you can control the example to start/stop sending packets by commands send, stop and exit:

[root@localhost dpdk-kcp]# ./dpdk-ikcp --socket-kcp --dst 1.1.1.1 --src 1.1.1.2 -v
cli > timestamp: 1550194806 recv 0 pkts in last 1000 milliseconds.
timestamp: 1550195807 recv 0 pkts in last 1000 milliseconds.
timestamp: 1550196808 recv 0 pkts in last 1000 milliseconds.

cli > support commands: send stop exit