[dpdk-dev] Appropriate DPDK data structures for TCP sockets

Matthew Hall mhall at mhcomputing.net
Sun Feb 22 01:38:45 CET 2015


Hello fellow stack hackers :) ,

I'm working on a simple server-side implementation of TCP on DPDK.

For this to work I need a good data structure to store some sockets.

The lookup key is like this:

struct ss_flow_key_s {
    uint8_t  sip[IPV6_ALEN];
    uint8_t  dip[IPV6_ALEN];
    uint16_t sport;
    uint16_t dport;
    uint8_t  protocol;
} __attribute__((packed));

typedef struct ss_flow_key_s ss_flow_key_t;

The socket itself is like this:

enum ss_tcp_state_e {
    SS_TCP_CLOSED   = 0, SS_TCP_LISTEN   = 1, SS_TCP_SYN_TX   = 2,
    SS_TCP_SYN_RX   = 3, SS_TCP_OPEN     = 4, SS_TCP_UNKNOWN  = -1,
};

typedef enum ss_tcp_state_e ss_tcp_state_t;

// RFC 793, RFC 1122
struct ss_tcp_socket_s {
    ss_flow_key_t  key;
    uint32_t id;
    rte_spinlock_t lock;

    ss_tcp_state_t state;

    uint32_t ticks_last;
    uint16_t rx_buf_offset;
    uint16_t mss;
    uint64_t rx_failures;
    uint8_t  rx_data[L4_TCP_WINDOW_SIZE * 2];
} __rte_cache_aligned;

So far I was using rte_hash, but it's single writer multi reader, which is eventually going to need some more complicated locking and probably run kind of slow. Also, I need some timer functions to delete dead sockets and so forth, and rte_hash doesn't have any iteration API.

So then I was trying to figure out if I needed to use a linked list for the iteration or if there is some other API I should use instead like rte_table_*. However the documentation of rte_table is kind of confusing so I wasn't sure if that was the right choice either.

Any advice?

Thanks,
Matthew.


More information about the dev mailing list