BIND9的架构与机制笔记1

　　BIND9采用的是事件驱动的机制来工作，而事件的源头则是IO，IO在linux使用的EPOLL的边缘触发模式。

　　本篇说的是epoll，BIND9如果创建了watcher线程（宏USE_WATCHER_THREAD控制），这里就讨论有线程的情况，实际上即使不创建

线程干的也都是一样的活。在lib/isc/socket.c中setup_watcher函数：（所有的代码都是截取的epoll下的片段，因为还有kqueue，devpoll，select等的实现代码，太多了）

#elif defined(USE_EPOLL)
    manager->nevents = ISC_SOCKET_MAXEVENTS;
    manager->events = isc_mem_get(mctx, sizeof(struct epoll_event) *
                      manager->nevents);
    if (manager->events == NULL)
        return (ISC_R_NOMEMORY);
    manager->epoll_fd = epoll_create(manager->nevents);
    if (manager->epoll_fd == -) {
        result = isc__errno2result(errno);
        isc__strerror(errno, strbuf, sizeof(strbuf));
        UNEXPECTED_ERROR(__FILE__, __LINE__,
                 "epoll_create %s: %s",
                 isc_msgcat_get(isc_msgcat, ISC_MSGSET_GENERAL,
                        ISC_MSG_FAILED, "failed"),
                 strbuf);
        isc_mem_put(mctx, manager->events,
                sizeof(struct epoll_event) * manager->nevents);
        return (result);
    }
#ifdef USE_WATCHER_THREAD
    result = watch_fd(manager, manager->pipe_fds[], SELECT_POKE_READ);
    if (result != ISC_R_SUCCESS) {
        close(manager->epoll_fd);
        isc_mem_put(mctx, manager->events,
                sizeof(struct epoll_event) * manager->nevents);
        return (result);
    }
#endif    /* USE_WATCHER_THREAD */

先是创建了要监视的最大socket fd数目（manager->nevents）对应的epoll_event结构体数组，然后调用epoll_create函数创建一个epoll fd，参数则是指定监视的socket fd

最大数目。我的内核版本是3.13，man一下epoll_create发现它是这样说的：epoll_create()  creates  an  epoll(7) instance.  Since Linux 2.6.8, thesize argument is ignored, but must be  greater  than  zero。这个函数在2.6.8内核以后就忽略参数size了，但是传递的参数值一定要大于0。后来找了一下资料，网上的高手的博客说的就很清楚了http://www.cnblogs.com/apprentice89/p/3234677.html。继续往下说，后面的watch_fd实在创建线程的情况下才有，就是将pipe_fds[0]这个管道描述符，也就是一个可读的流，而上述的socket fd都是可以归为流。watch_fd的实现代码：

#elif defined(USE_EPOLL)
        struct epoll_event event;

        if (msg == SELECT_POKE_READ)
                event.events = EPOLLIN;
        else
                event.events = EPOLLOUT;
        memset(&event.data, , sizeof(event.data));
        event.data.fd = fd;
        if (epoll_ctl(manager->epoll_fd, EPOLL_CTL_ADD, fd, &event) == - &&
            errno != EEXIST) {
                result = isc__errno2result(errno);
        }

        return (result);

这是将pipe_fds[0]加入epoll_fd的监听队列，EPOLL_CTL_ADD是操作类型，注册该fd到epoll_fd上。这个管道的目的是接收管理该线程的消息，比如线程退出。

那么进入线程看：

static isc_threadresult_t
watcher(void *uap) {
    isc__socketmgr_t *manager = uap;
    isc_boolean_t done;
    int ctlfd;
    int cc;
#ifdef USE_KQUEUE
    const char *fnname = "kevent()";
#elif defined (USE_EPOLL)
    const char *fnname = "epoll_wait()";
#elif defined(USE_DEVPOLL)
    const char *fnname = "ioctl(DP_POLL)";
    struct dvpoll dvp;
#elif defined (USE_SELECT)
    const char *fnname = "select()";
    int maxfd;
#endif
    char strbuf[ISC_STRERRORSIZE];
#ifdef ISC_SOCKET_USE_POLLWATCH
    pollstate_t pollstate = poll_idle;
#endif

    /*
     * Get the control fd here.  This will never change.
     */
    ctlfd = manager->pipe_fds[];
    done = ISC_FALSE;
    while (!done) {
        do {
#ifdef USE_KQUEUE
            cc = kevent(manager->kqueue_fd, NULL, ,
                    manager->events, manager->nevents, NULL);
#elif defined(USE_EPOLL)
            cc = epoll_wait(manager->epoll_fd, manager->events,
                    manager->nevents, -);
#elif defined(USE_DEVPOLL)
            dvp.dp_fds = manager->events;
            dvp.dp_nfds = manager->nevents;
#ifndef ISC_SOCKET_USE_POLLWATCH
            dvp.dp_timeout = -;
#else
            if (pollstate == poll_idle)
                dvp.dp_timeout = -;
            else
                dvp.dp_timeout = ISC_SOCKET_POLLWATCH_TIMEOUT;
#endif    /* ISC_SOCKET_USE_POLLWATCH */
            cc = ioctl(manager->devpoll_fd, DP_POLL, &dvp);
#elif defined(USE_SELECT)
            LOCK(&manager->lock);
            memcpy(manager->read_fds_copy, manager->read_fds,
                   manager->fd_bufsize);
            memcpy(manager->write_fds_copy, manager->write_fds,
                   manager->fd_bufsize);
            maxfd = manager->maxfd + ;
            UNLOCK(&manager->lock);

            cc = select(maxfd, manager->read_fds_copy,
                    manager->write_fds_copy, NULL, NULL);
#endif    /* USE_KQUEUE */

            if (cc <  && !SOFT_ERROR(errno)) {
                isc__strerror(errno, strbuf, sizeof(strbuf));
                FATAL_ERROR(__FILE__, __LINE__,
                        "%s %s: %s", fnname,
                        isc_msgcat_get(isc_msgcat,
                               ISC_MSGSET_GENERAL,
                               ISC_MSG_FAILED,
                               "failed"), strbuf);
            }

#if defined(USE_DEVPOLL) && defined(ISC_SOCKET_USE_POLLWATCH)
            if (cc == ) {
                if (pollstate == poll_active)
                    pollstate = poll_checking;
                else if (pollstate == poll_checking)
                    pollstate = poll_idle;
            } else if (cc > ) {
                if (pollstate == poll_checking) {
                    /*
                     * XXX: We'd like to use a more
                     * verbose log level as it's actually an
                     * unexpected event, but the kernel bug
                     * reportedly happens pretty frequently
                     * (and it can also be a false positive)
                     * so it would be just too noisy.
                     */
                    manager_log(manager,
                            ISC_LOGCATEGORY_GENERAL,
                            ISC_LOGMODULE_SOCKET,
                            ISC_LOG_DEBUG(),
                            "unexpected POLL timeout");
                }
                pollstate = poll_active;
            }
#endif
        } while (cc < );

#if defined(USE_KQUEUE) || defined (USE_EPOLL) || defined (USE_DEVPOLL)
        done = process_fds(manager, manager->events, cc);
#elif defined(USE_SELECT)
        process_fds(manager, maxfd, manager->read_fds_copy,
                manager->write_fds_copy);

        /*
         * Process reads on internal, control fd.
         */
        if (FD_ISSET(ctlfd, manager->read_fds_copy))
            done = process_ctlfd(manager);
#endif
    }

    manager_log(manager, TRACE, "%s",
            isc_msgcat_get(isc_msgcat, ISC_MSGSET_GENERAL,
                   ISC_MSG_EXITING, "watcher exiting"));

    return ((isc_threadresult_t));
}

无限循环，epoll_wait当监听的epoll_fd队列上有IO事件发生时，将对应的socket fd和事件放入events数组中，并且将这些注册在epoll_fd上的socket fd对应事件清空。

process_fds遍历数组，找到对应的socket fd，并判断该fd是不是线程控制管道，如果是则会在执行完其他socket fd上的对应事件后再处理管道中的控制消息。

static isc_boolean_t
process_fds(isc__socketmgr_t *manager, struct epoll_event *events, int nevents)
{
    int i;
    isc_boolean_t done = ISC_FALSE;
#ifdef USE_WATCHER_THREAD
    isc_boolean_t have_ctlevent = ISC_FALSE;
#endif

    if (nevents == manager->nevents) {
        manager_log(manager, ISC_LOGCATEGORY_GENERAL,
                ISC_LOGMODULE_SOCKET, ISC_LOG_INFO,
                "maximum number of FD events (%d) received",
                nevents);
    }

    for (i = ; i < nevents; i++) {
        REQUIRE(events[i].data.fd < (int)manager->maxsocks);
#ifdef USE_WATCHER_THREAD
        if (events[i].data.fd == manager->pipe_fds[]) {
            have_ctlevent = ISC_TRUE;
            continue;
        }
#endif
        if ((events[i].events & EPOLLERR) !=  ||
            (events[i].events & EPOLLHUP) != ) {
            /*
             * epoll does not set IN/OUT bits on an erroneous
             * condition, so we need to try both anyway.  This is a
             * bit inefficient, but should be okay for such rare
             * events.  Note also that the read or write attempt
             * won't block because we use non-blocking sockets.
             */
            events[i].events |= (EPOLLIN | EPOLLOUT);
        }
        process_fd(manager, events[i].data.fd,
               (events[i].events & EPOLLIN) != ,
               (events[i].events & EPOLLOUT) != );
    }

#ifdef USE_WATCHER_THREAD
    if (have_ctlevent)
        done = process_ctlfd(manager);
#endif

    return (done);
}

待续