在init进程的main函数中有调用sigchld_handler_init() 设置SIGCHLD signal。这里设置SIGCHLD的处理函数时，在sa_flags中有加SA_NOCLDSTOP flag，这个flag表示只有当子进程终止时父进程才接受这个signal，当子进程在暂停或者继续运行的状态时是不会收到这个signal的。
void sigchld_handler_init() {// Create a signalling mechanism for SIGCHLD.int s[2];if (socketpair(AF_UNIX, SOCK_STREAM | SOCK_NONBLOCK | SOCK_CLOEXEC, 0, s) == -1) {PLOG(FATAL) << "socketpair failed in sigchld_handler_init";}signal_write_fd = s[0];signal_read_fd = s[1];// Write to signal_write_fd if we catch SIGCHLD.struct sigaction act;memset(&act, 0, sizeof(act));act.sa_handler = SIGCHLD_handler;act.sa_flags = SA_NOCLDSTOP;sigaction(SIGCHLD, &act, 0);ReapAnyOutstandingChildren();register_epoll_handler(signal_read_fd, handle_signal);}
1. 注册的SIGCHLD处理函数是SIGCHLD_handler。
2. 调用register_epoll_handler()
init.cpp
void register_epoll_handler(int fd, void (*fn)()) {epoll_event ev;ev.events = EPOLLIN;ev.data.ptr = reinterpret_cast<void*>(fn);if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, fd, &ev) == -1) {PLOG(ERROR) << "epoll_ctl failed";}}
1. 在子进程退出时，init进程会收到SIGCHLD signal，然后调用SIGCHLD_handler()进行处理。这个函数只是简单地调用write()
2. 由于上面有进行write，所以在init的main函数中的epoll_wait()会返回，然后调用通过register_epoll_handler()注册的函数，即hand_signal()
epoll_event ev;int nr = TEMP_FAILURE_RETRY(epoll_wait(epoll_fd, &ev, 1, epoll_timeout_ms));if (nr == -1) {PLOG(ERROR) << "epoll_wait failed";} else if (nr == 1) {((void (*)()) ev.data.ptr)();}
在hand_signal()中，首先调用read()将数据读出，然后会调用到ReapOneProcess()。在这个函数中，调用waitpid()等待子进程退出，waitpid()会拿到一个siginfo，这个siginfo中有退出的子进程的pid。
然后根据这个pid从ServiceList中找到这个退出的子进程的Service对象；
然后调用这个Service对象的Reap()。
在Reap函数中，首先调用KillProcessGroup(SIGKILL)。这个函数会调到system/core/libprocessgroup/processgroup.cpp中的killProcessGroup()。这个调用的目的是kill掉退出的子进程的所有子进程。
void Service::KillProcessGroup(int signal) {// If we've already seen a successful result from killProcessGroup*(), then we have removed// the cgroup already and calling these functions a second time will simply result in an error.// This is true regardless of which signal was sent.// These functions handle their own logging, so no additional logging is needed.if (!process_cgroup_empty_) {LOG(INFO) << "Sending signal " << signal << " to service '" << name_ << "' (pid " << pid_<< ") process group...";int r;if (signal == SIGTERM) {r = killProcessGroupOnce(uid_, pid_, signal);} else {r = killProcessGroup(uid_, pid_, signal);}if (r == 0) process_cgroup_empty_ = true;}}
然后调用UnsetExec()将is_exec_service_running_设置为false并将SVC_EXEC flag清掉；
然后会将flags_中的SVC_RESTARTING flag设置上；
然后调用onrestart_.ExecuteAllCommands()
void Service::Reap(const siginfo_t& siginfo) {if (!(flags_ & SVC_ONESHOT) || (flags_ & SVC_RESTART)) {KillProcessGroup(SIGKILL);}// Remove any descriptor resources we may have created.std::for_each(descriptors_.begin(), descriptors_.end(),std::bind(&DescriptorInfo::Clean, std::placeholders::_1));for (const auto& f : reap_callbacks_) {f(siginfo);}if (flags_ & SVC_EXEC) UnSetExec();if (flags_ & SVC_TEMPORARY) return;pid_ = 0;flags_ &= (~SVC_RUNNING);start_order_ = 0;// Oneshot processes go into the disabled state on exit,// except when manually restarted.if ((flags_ & SVC_ONESHOT) && !(flags_ & SVC_RESTART)) {flags_ |= SVC_DISABLED;}// Disabled and reset processes do not get restarted automatically.if (flags_ & (SVC_DISABLED | SVC_RESET)) {NotifyStateChange("stopped");return;}// If we crash > 4 times in 4 minutes, reboot into recovery.boot_clock::time_point now = boot_clock::now();if ((flags_ & SVC_CRITICAL) && !(flags_ & SVC_RESTART)) {if (now < time_crashed_ + 4min) {if (++crash_count_ > 4) {LOG(FATAL) << "critical process '" << name_ << "' exited 4 times in 4 minutes";}} else {time_crashed_ = now;crash_count_ = 1;}}flags_ &= (~SVC_RESTART); flags_ |= SVC_RESTARTING;// Execute all onrestart commands for this service.onrestart_.ExecuteAllCommands();NotifyStateChange("restarting");return;}
然后在init的main函数的死循环中，会调用Service::is_exec_service_running()函数，这个函数是直接return Service class的is_exec_service_running_ static bool变量（注意这个变量是static变量）。
因为上面有将is_exec_service_running_变量设置为false，所以这个if条件成立，所以会调用RestartProcesses()。在这个函数中，会从ServiceList中取出每一个Service对象，然后判断其flags_变量中是否有SVC_RESTARTING flag，如果有，最终会调到这个service的Start()去再次启动这个service。
