Thursday, November 22, 2012

OpenvSwitch 代码分析(二)


本小节分析vswitchd模块,该模块作为管理底层datapath的重要模块,实现了openflow的主要逻辑,以及对交换机的具体管理和除简单转发外的其他逻辑。可见,该模块十分重要,主要生成ovs-vswitchd文件,其中主文件为vswitchd/ovs-vswitchd.c

整体分析

Vswitchd模块主要包括bridgeofproto等子模块。作为主要逻辑实现模块,负责解析和执行其他各个openflow命令。

bridge模块

负责所管理的所有datapath,对外的接口很简单,包括
void bridge_init(const char *remote);
void bridge_exit(void);

void bridge_run(void);
void bridge_run_fast(void);
void bridge_wait(void);

void bridge_get_memory_usage(struct simap *usage);
数据结构主要在bridge.c中定义了bridge结构,定义为
struct bridge {
    struct hmap_node node;      /* In 'all_bridges'. */
    char *name;                 /* User-specified arbitrary name. */
    char *type;                 /* Datapath type. */
    uint8_t ea[ETH_ADDR_LEN];   /* Bridge Ethernet Address. */
    uint8_t default_ea[ETH_ADDR_LEN]; /* Default MAC. */
    const struct ovsrec_bridge *cfg;

    /* OpenFlow switch processing. */
    struct ofproto *ofproto;    /* OpenFlow switch. */

    /* Bridge ports. */
    struct hmap ports;          /* "struct port"s indexed by name. */
    struct hmap ifaces;         /* "struct iface"s indexed by ofp_port. */
    struct hmap iface_by_name;  /* "struct iface"s indexed by name. */

    struct list ofpp_garbage;   /* "struct ofpp_garbage" slated for removal. */
    struct hmap if_cfg_todo;    /* "struct if_cfg"s slated for creation.
                                   Indexed on 'cfg->name'. */

    /* Port mirroring. */
    struct hmap mirrors;        /* "struct mirror" indexed by UUID. */

    /* Synthetic local port if necessary. */
    struct ovsrec_port synth_local_port;
    struct ovsrec_interface synth_local_iface;
    struct ovsrec_interface *synth_local_ifacep;
};
其中,最重要的是ofproto指针,指向一个openflow switch,负责进行openflow switch的所有处理。实际上,vswitchd的主要功能就是不断检测并调用所有bridge上的ofproto,执行其上的处理函数。

ofproto

类型定义在ofproto/ofproto-provider.h中。
struct ofproto {
    struct hmap_node hmap_node; /* In global 'all_ofprotos' hmap. */
    const struct ofproto_class *ofproto_class;
    char *type;                 /* Datapath type. */
    char *name;                 /* Datapath name. */

    /* Settings. */
    uint64_t fallback_dpid;     /* Datapath ID if no better choice found. */
    uint64_t datapath_id;       /* Datapath ID. */
    unsigned flow_eviction_threshold; /* Threshold at which to begin flow
                                       * table eviction. Only affects the
                                       * ofproto-dpif implementation */
    bool forward_bpdu;          /* Option to allow forwarding of BPDU frames
                                 * when NORMAL action is invoked. */
    char *mfr_desc;             /* Manufacturer. */
    char *hw_desc;              /* Hardware. */
    char *sw_desc;              /* Software version. */
    char *serial_desc;          /* Serial number. */
    char *dp_desc;              /* Datapath description. */
    enum ofp_config_flags frag_handling; /* One of OFPC_*.  */

    /* Datapath. */
    struct hmap ports;          /* Contains "struct ofport"s. */
    struct shash port_by_name;

    /* Flow tables. */
    struct oftable *tables;
    int n_tables;

    /* OpenFlow connections. */
    struct connmgr *connmgr;

    /* Flow table operation tracking. */
    int state;                  /* Internal state. */
    struct list pending;        /* List of "struct ofopgroup"s. */
    unsigned int n_pending;     /* list_size(&pending). */
    struct hmap deletions;      /* All OFOPERATION_DELETE "ofoperation"s. */

    /* Flow table operation logging. */
    int n_add, n_delete, n_modify; /* Number of unreported ops of each kind. */
    long long int first_op, last_op; /* Range of times for unreported ops. */
    long long int next_op_report;    /* Time to report ops, or LLONG_MAX. */
    long long int op_backoff;        /* Earliest time to report ops again. */

    /* Linux VLAN device support (e.g. "eth0.10" for VLAN 10.)
     *
     * This is deprecated.  It is only for compatibility with broken device
     * drivers in old versions of Linux that do not properly support VLANs when
     * VLAN devices are not used.  When broken device drivers are no longer in
     * widespread use, we will delete these interfaces. */
    unsigned long int *vlan_bitmap; /* 4096-bit bitmap of in-use VLANs. */
    bool vlans_changed;             /* True if new VLANs are in use. */
    int min_mtu;                    /* Current MTU of non-internal ports. */
};
其中最关键的是ofproto_class,是ofproto交换机的具体实现,定义了对于of协议的处理(包括runrun_fast函数,前者处理更为全面,调用了后者)。处理函数的实现在ofproto/ofproto-dpif.c中。

run_fast()

位于ofproto-dpif.c中。
分析run_fast函数,主要完成了两个需要周期性及时完成的事情。
首先对各个port上调用port_run_fast,检查是否要发送连续性检查的网包消息(CCM,参考IEEE 802.1aq),如果是,则发出。
HMAP_FOR_EACH (ofport, up.hmap_node, &ofproto->up.ports) {
        port_run_fast(ofport);
    }
然后,检查是否有upcall,对所有的来自datapathupcall进行处理。
while (work < FLOW_MISS_MAX_BATCH) {
        int retval = handle_upcalls(ofproto, FLOW_MISS_MAX_BATCH - work);
        if (retval <= 0) {
            return -retval;
        }
        work += retval;
    }

handle_upcalls()

位于ofproto-dpif.c中。
该函数从对应的dpif中获取到upcalls后,对upcalls进行类型检查。对于SFLOW_UPCALLBAD_UPCALL,进行对应处理后释放存有upcall消息的buf,而对于MISS_UPCALL类型,则调用handle_miss_upcalls进行后续的处理。
其中,upcall的类型为dpif_upcalllib/dpif.h),定义为
/* A packet passed up from the datapath to userspace.
 *
 * If 'key' or 'actions' is nonnull, then it points into data owned by
 * 'packet', so their memory cannot be freed separately.  (This is hardly a
 * great way to do things but it works out OK for the dpif providers and
 * clients that exist so far.)
 */
struct dpif_upcall {
    /* All types. */
    enum dpif_upcall_type type;
    struct ofpbuf *packet;      /* Packet data. */
    struct nlattr *key;         /* Flow key. */
    size_t key_len;             /* Length of 'key' in bytes. */

    /* DPIF_UC_ACTION only. */
    uint64_t userdata;          /* Argument to OVS_ACTION_ATTR_USERSPACE. */
};

handle_miss_upcalls ()

位于ofproto-dpif.c中。
该函数从upcall中提取相关的流信息,把属于同一个key的网包放到一起,最后放进todo list中。最后检查todo list中的每个元素,调用handle_flow_miss()进行处理。
HMAP_FOR_EACH (miss, hmap_node, &todo) {
        handle_flow_miss(ofproto, miss, flow_miss_ops, &n_ops);
}
处理完毕后,调用dpif_operate()(位于vswitchd/dpif.c)执行查找到的行动。
for (i = 0; i < n_ops; i++) {
        dpif_ops[i] = &flow_miss_ops[i].dpif_op;
    }
    dpif_operate(ofproto->dpif, dpif_ops, n_ops);

handle_miss_upcall ()

位于ofproto/ofproto-dpif.c中。
该函数处理给定的某个miss_upcall。首先,先判断是否发生了精确匹配,如果发生了,则直接按照匹配结果调用handle_flow_miss_with_facet();如果没有精确匹配结果,则调用handle_flow_miss_without_facet()

unixctl_server相关

主循环中调用了unixctl_server_run()函数。该函数首先获取到远端server的连接,然后,执行连接中的命令,代码为。
LIST_FOR_EACH_SAFE (conn, next, node, &server->conns) {
        int error = run_connection(conn);
        if (error && error != EAGAIN) {
            kill_connection(conn);
        }
    }

ovs-vswitchd.c

主文件,其中main()为入口主函数,执行一系列的初始化,并配置各个队列,最后是主循环,进行任务处理。分析主要代码如下
Int main(int argc, char *argv[])
{
    char *unixctl_path = NULL;
    struct unixctl_server *unixctl;
    struct signal *sighup;
    char *remote;
    bool exiting;
    int retval;

    proctitle_init(argc, argv); //backup orignal argvs
    set_program_name(argv[0]);
    stress_init_command(); //register stress cmds to the commands
    remote = parse_options(argc, argv, &unixctl_path);
    signal(SIGPIPE, SIG_IGN); //ignore the pipe read end signal
    sighup = signal_register(SIGHUP); //register the SIGHUP signal handler
    process_init(); //create notification pipe and register signal for child process exit
    ovsrec_init(); //todo: make clear here
    daemonize_start(); //daemonize the process

    if (want_mlockall) {
#ifdef HAVE_MLOCKALL
        if (mlockall(MCL_CURRENT | MCL_FUTURE)) {
            VLOG_ERR("mlockall failed: %s", strerror(errno));
        }
#else
        VLOG_ERR("mlockall not supported on this system");
#endif
    }
    worker_start(); //start a worker subprocess, call worker_main (receive data and process)
    retval = unixctl_server_create(unixctl_path, &unixctl);//create a unix domain socket
    if (retval) {
        exit(EXIT_FAILURE);
    }
    unixctl_command_register("exit", "", 0, 0, ovs_vswitchd_exit, &exiting);

    bridge_init(remote);//ini the bridge, configure from the ovsdb server, register ctrl commands
    free(remote);
    exiting = false;
    while (!exiting) {
        worker_run(); //reply with the worker subprocess
        if (signal_poll(sighup)) {
            vlog_reopen_log_file();
        }
        memory_run();//monitor the memory
        if (memory_should_report()) {
            struct simap usage;
            simap_init(&usage);
            bridge_get_memory_usage(&usage);
            memory_report(&usage);
            simap_destroy(&usage);
        }
        bridge_run_fast(); //check each bridge and run it's handler
        bridge_run(); //main process part, process of pkts
        bridge_run_fast();
        unixctl_server_run(unixctl);
        netdev_run(); //run periodic functions by all network devices.
        worker_wait();
        signal_wait(sighup);
        memory_wait();
        bridge_wait();
        unixctl_server_wait(unixctl);
        netdev_wait();
        if (exiting) {
            poll_immediate_wake();
        }
        poll_block();
    }
    bridge_exit();
    unixctl_server_destroy(unixctl);
    signal_unregister(sighup);
    return 0;
}

proctitle_init(argc, argv)

复制出输入的参数列表到新的存储中,让argv指向这块内存,主要是为了后面的proctitle_set()函数(在daemonize_start()->monitor_daemon()中调用,可能修改原argv存储)做准备。

    set_program_name(argv[0])

设置程序名称、版本、编译日期等信息。

    stress_init_command()

注册stress 相关命令(listsetenabledisable)到commands结构。

    remote = parse_options(argc, argv, &unixctl_path)

解析参数,其中unixctl_path存储unixctrl域的sock名,作为接受外部控制命令的渠道;而remote存储连接到ovsdb的信息,即连接到配置数据库的sock名。

    signal(SIGPIPE, SIG_IGN)

忽略pipe读结束的信号。

    sighup = signal_register(SIGHUP)

注册对 SIGHUP信号(终端挂起)的处理函数。处理函数为写到fds[1]中空字符。

    process_init()

注册对SIGCHLD信号(子进程结束)的处理函数。处理函数为执行all_process上的所有进程。

    ovsrec_init()

数据表结构初始化。包括13张数据表。表的具体结构请参考ovsdb的相关文档。

daemonize_start()

让进程变成守护程序。

worker_start()

开启一个worker子进程。子进程与主进程交互数据。

unixctl_server_create(unixctl_path, &unixctl)

创建一个unixctl server(存放在unixctl),并监听在unixctl_path指定的punix路径。

unixctl_command_register("exit", "", 0, 0, vs_vswitchd_exit, &exiting)

注册unixctl命令。

    bridge_init(remote)

remote数据库获取配置信息,并初始化bridge

主循环

exiting = false;
    while (!exiting) {
        worker_run(); //reply with the worker subprocess
        if (signal_poll(sighup)) {
            vlog_reopen_log_file();
        }
        memory_run();//monitor the memory
        if (memory_should_report()) {
            struct simap usage;

            simap_init(&usage);
            bridge_get_memory_usage(&usage);
            memory_report(&usage);
            simap_destroy(&usage);
        }
        bridge_run_fast(); //check each bridge and run it's handler
        bridge_run(); //main process part, process of pkts
        bridge_run_fast();
        unixctl_server_run(unixctl);
        netdev_run(); //run periodic functions by all network devices.

        worker_wait();
        signal_wait(sighup);
        memory_wait();
        bridge_wait();
        unixctl_server_wait(unixctl);
        netdev_wait();
        if (exiting) {
            poll_immediate_wake();
        }
        poll_block();
}

worker_run()

执行从worker子进程中获取的RPC reply,执行其中的cb_reply回调函数。主要过程为
rxbuf_run(&client_rx, client_sock, sizeof(struct worker_reply));
reply->reply_cb(&client_rx.payload, client_rx.fds,  client_rx.n_fds, reply->reply_aux);

bridge_run_fast()

执行在all_bridge上的每个bridgeofproto上的run_fast。主要是监听和处理来自datapathupcall,主要过程为
HMAP_FOR_EACH (br, node, &all_bridges) {
        ofproto_run_fast(br->ofproto);
    }

bridge_run()

主要的对网包进行慢速处理过程。包括完成必要的配置更新(在配置更新中会从数据库读入配置信息,生成必要的bridgedp等数据结构),以及执行在all_bridge上的每个bridgeofproto上的run(),并做相应的信息统计。
其中,ofproto上的run()主要依次调用了如下函数:
调用dpif_run()处理所有注册的netlink notifier的汇报事件。
调用run_fast()处理常见的周期性事件,包括对upcalls的处理等。
可选调用netflow_run()sflow_run(),进行对netflowsflow的支持
可选调用port_run()进行发送CCM
可选调用bundle_run()处理LACPbonding等杂项。
可选调用stp_run()进行STP支持。
mac_learning_run()获取超时的mac entry,并将其删除掉。
可选调用governor_run()进行限速处理。

unixctl_server_run(unixctl)

unixctl指定的server中获取数据,并执行对应的配置命令。主要过程为
struct unixctl_conn *conn = xzalloc(sizeof *conn);
            list_push_back(&server->conns, &conn->node);
            conn->rpc = jsonrpc_open(stream);

netdev_run()

执行在netdev_classes上定义的每个netdev_class实体,调用它们的run()。主要过程为
SHASH_FOR_EACH(node, &netdev_classes) {
        const struct netdev_class *netdev_class = node->data;
        if (netdev_class->run) {
            netdev_class->run();
        }
}

循环等待事件处理

包括wokersignalmemorybridgeunixctl_servernetdev等事件,被poll_fd_wait()注册。

poll_block(void)

阻塞,直到之前被poll_fd_wait()注册过的事件发生,或者等待时间超过poll_timer_wait()注册的最短时间。

清理工作

退出bridge,关闭unixctl 连接,取消对sighup信号的处理注册。
bridge_exit();
unixctl_server_destroy(unixctl);
signal_unregister(sighup);

通用类型

基础宏定义

首先分析下几个常见的基础宏。
CONTAINER_OF宏:返回拥有某个给定memberstruct结构的起始地址。其中,struct为所定义的数据结构,其中有一个变量名字为memberpointer为指向member变量的一个指针。该宏返回整个数据结构的起始地址。定义为
#define CONTAINER_OF(POINTER, STRUCT, MEMBER)                           \
        ((STRUCT *) (void *) ((char *) (POINTER) - offsetof (STRUCT, MEMBER)))

OBJECT_CONTAINING宏:返回含有某个给定member的对象object的地址。
#define OBJECT_CONTAINING(POINTER, OBJECT, MEMBER)                      \
    ((OVS_TYPEOF(OBJECT)) (void *)                                      \
     ((char *) (POINTER) - OBJECT_OFFSETOF(OBJECT, MEMBER)))

ASSIGN_CONTAINER宏:返回含有某个给定member的对象object1
#define ASSIGN_CONTAINER(OBJECT, POINTER, MEMBER) \
    ((OBJECT) = OBJECT_CONTAINING(POINTER, OBJECT, MEMBER), 1)

普通列表

ovs代码中大量使用了列表的结构。列表的声明在lib/list.h中。列表的抽象结构用户不必关心,用户只需要维护好自己关心的节点的数据结构即可。
使用一个list struct list L = LIST_INITIALIZER(&L)
正向遍历:
#define LIST_FOR_EACH(ITER, MEMBER, LIST)                               \
    for (ASSIGN_CONTAINER(ITER, (LIST)->next, MEMBER);                  \
         &(ITER)->MEMBER != (LIST);                                     \
         ASSIGN_CONTAINER(ITER, (ITER)->MEMBER.next, MEMBER))
逆向遍历
#define LIST_FOR_EACH_REVERSE(ITER, MEMBER, LIST)                       \
    for (ASSIGN_CONTAINER(ITER, (LIST)->prev, MEMBER);                  \
         &(ITER)->MEMBER != (LIST);                                     \
         ASSIGN_CONTAINER(ITER, (ITER)->MEMBER.prev, MEMBER))
安全遍历
#define LIST_FOR_EACH_SAFE(ITER, NEXT, MEMBER, LIST)            \
    for (ASSIGN_CONTAINER(ITER, (LIST)->next, MEMBER);          \
         (&(ITER)->MEMBER != (LIST)                             \
          ? ASSIGN_CONTAINER(NEXT, (ITER)->MEMBER.next, MEMBER) \
          : 0);                                                 \
         (ITER) = (NEXT))

Hash列表

列表的声明在lib/hmap.h中。
Hmap列表,包含两个指针,指向其中含有的节点,定义为
/* A hash map. */
struct hmap {
    struct hmap_node **buckets; /* Must point to 'one' iff 'mask' == 0. */
    struct hmap_node *one;
    size_t mask;
    size_t n;
};
hmap_node类型,包括一个hash值和一个后继指针,定义为
struct hmap_node {
    size_t hash;                /* Hash value. */
    struct hmap_node *next;     /* Next in linked list. */
};
hmap的列表遍历似乎通过如下宏来实现的
#define HMAP_FOR_EACH(NODE, MEMBER, HMAP)                               \
    for (ASSIGN_CONTAINER(NODE, hmap_first(HMAP), MEMBER);              \
         &(NODE)->MEMBER != NULL;                                       \
         ASSIGN_CONTAINER(NODE, hmap_next(HMAP, &(NODE)->MEMBER), MEMBER))
其中ASSIGN_CONTAINER有三个参数,分别是一个输出指针,一个输入指针和一个成员。输入指针指向该成员,而输出指针获取指向包含有该成员的类型的地址。
所以遍历宏实际上,遍历了hmap的每个节点,并将包含该节点的数据结构逐个返回。

ofproto

ofproto_class

log消息

#define VLOG_FATAL(...) vlog_fatal(THIS_MODULE, __VA_ARGS__)
#define VLOG_ABORT(...) vlog_abort(THIS_MODULE, __VA_ARGS__)
#define VLOG_EMER(...) VLOG(VLL_EMER, __VA_ARGS__)
#define VLOG_ERR(...) VLOG(VLL_ERR, __VA_ARGS__)
#define VLOG_WARN(...) VLOG(VLL_WARN, __VA_ARGS__)
#define VLOG_INFO(...) VLOG(VLL_INFO, __VA_ARGS__)
#define VLOG_DBG(...) VLOG(VLL_DBG, __VA_ARGS__)


No comments:

Post a Comment