在工做中,咱们常常须要重启PHP-FPM,那么这个重启过程都发生了那些事情呢?让咱们从PHP源码中一探究竟吧。php
运行环境: Mac 10.14.2 + PHP 7.3.7segmentfault
信号在fpm的重启中扮演着重要的角色。那什么是信号呢?异步
信号是由用户、系统或者进程发送给目标进程的信息,以通知目标进程某个状态的改变或系统异常。Linux信号可由以下条件产生:socket
- 对于前台进程,用户能够经过输入特殊的终端字符来给它发送信号。
- 系统异常。好比浮点异常和非法内存段访问。
- 系统状态变化。好比 alarm 定时器到期将引发 SIGALARM 信号。
- 运行 kill 命令或调用 kill 函数
在PHP-FPM中,用户经过kill
命令来重启fpm,master进程也是经过kill()
函数向worker进程发送信号来结束进程。fpm的重启分为优雅重启(kill -SIGUSR2
)和强制重启(kill -SIGTERM
)两种,下面是以优雅重启为例,master进程将收到SIGUSR2
信号。ide
master进程信号初始化函数fpm_signals_init_main()
主要作了两件事情:函数
经过socketpair()
来建立这一对双全工的unix_socket,其中sp[0]
的可读事件在fpm_event_loop()
中被注册到事件队列中,其回调函数为fpm_got_signal()
,这样往sp[1]
写入数据时将触发sp[0]
的可读事件回调。对这俩unix_socket还有两个操做:php-fpm
fcntl(fd, F_SETFL, old_flags|O_NONBLOCK)
,这样当fd不可读或不可写的时候,read()
、write()
不会阻塞,而是直接返回-1,errno设为EAGAIN。fcntl(fd, F_SETFD, FD_CLOEXEC)
,这样当进程调用exec()
族函数前会关闭该fd。这么作是为了防止文件描述符的泄露,由于调用exec()
族函数会用新程序替换掉当前进程执行的程序,进程的正文、数据、堆和栈段都会被替换,这就致使原先保存文件描述符的变量不存在了,也就没法关闭“老进程“的fd,致使文件描述符泄露。注册的信号有SIGTERM
、SIGINT
、SIGUSR1
、SIGUSR2
、SIGCHLD
、SIGQUIT
六种。oop
int fpm_signals_init_main() /* {{{ */ { struct sigaction act; if (0 > socketpair(AF_UNIX, SOCK_STREAM, 0, sp)) { zlog(ZLOG_SYSERROR, "failed to init signals: socketpair()"); return -1; } if (0 > fd_set_blocked(sp[0], 0) || 0 > fd_set_blocked(sp[1], 0)) { zlog(ZLOG_SYSERROR, "failed to init signals: fd_set_blocked()"); return -1; } if (0 > fcntl(sp[0], F_SETFD, FD_CLOEXEC) || 0 > fcntl(sp[1], F_SETFD, FD_CLOEXEC)) { zlog(ZLOG_SYSERROR, "falied to init signals: fcntl(F_SETFD, FD_CLOEXEC)"); return -1; } memset(&act, 0, sizeof(act)); act.sa_handler = sig_handler; sigfillset(&act.sa_mask); if (0 > sigaction(SIGTERM, &act, 0) || 0 > sigaction(SIGINT, &act, 0) || 0 > sigaction(SIGUSR1, &act, 0) || 0 > sigaction(SIGUSR2, &act, 0) || 0 > sigaction(SIGCHLD, &act, 0) || 0 > sigaction(SIGQUIT, &act, 0)) { zlog(ZLOG_SYSERROR, "failed to init signals: sigaction()"); return -1; } return 0; }
worker进程信号初始化函数fpm_signals_init_child()
主要作了三件事情:ui
这对unix_socket继承自master进程,worker进程用不到它们。.net
sig_soft_quit()
,sa_flags
变量设为SA_RESTART
表示信号处理函数返回后从新调用被中断的系统调用,这样worker进程正在处理中的事情不会受到影响。SIG_DFL
,即采用默认行为。调用zend_signal_init()
,这个不展开讲了。
int fpm_signals_init_child() /* {{{ */ { struct sigaction act, act_dfl; memset(&act, 0, sizeof(act)); memset(&act_dfl, 0, sizeof(act_dfl)); act.sa_handler = &sig_soft_quit; act.sa_flags |= SA_RESTART; act_dfl.sa_handler = SIG_DFL; close(sp[0]); close(sp[1]); if (0 > sigaction(SIGTERM, &act_dfl, 0) || 0 > sigaction(SIGINT, &act_dfl, 0) || 0 > sigaction(SIGUSR1, &act_dfl, 0) || 0 > sigaction(SIGUSR2, &act_dfl, 0) || 0 > sigaction(SIGCHLD, &act_dfl, 0) || 0 > sigaction(SIGQUIT, &act, 0)) { zlog(ZLOG_SYSERROR, "failed to init child signals: sigaction()"); return -1; } zend_signal_init(); return 0; }
master进程收到SIGUSR2
信号后将回调sig_handler()
进行信号处理。咱们能够看到SIGUSR2
被映射为2
,并写入到 sp[1]
。
static void sig_handler(int signo) /* {{{ */ { static const char sig_chars[NSIG + 1] = { [SIGTERM] = 'T', [SIGINT] = 'I', [SIGUSR1] = '1', [SIGUSR2] = '2', [SIGQUIT] = 'Q', [SIGCHLD] = 'C' }; char s; int saved_errno; if (fpm_globals.parent_pid != getpid()) { /* prevent a signal race condition when child process have not set up it's own signal handler yet */ return; } saved_errno = errno; s = sig_chars[signo]; zend_quiet_write(sp[1], &s, sizeof(s)); //实际调用write() errno = saved_errno; }
当往sp[1]
写入数据后,sp[0]
变为可读,触发事件回调fpm_got_signal()
。从sp[0]
读取到写入的数据 2
,以后调用fpm_pctl()
来进行重启操做。
static void fpm_got_signal(struct fpm_event_s *ev, short which, void *arg) /* {{{ */ { char c; int res, ret; int fd = ev->fd; do { do { res = read(fd, &c, 1); } while (res == -1 && errno == EINTR); if (res <= 0) { if (res < 0 && errno != EAGAIN && errno != EWOULDBLOCK) { zlog(ZLOG_SYSERROR, "unable to read from the signal pipe"); } return; } switch (c) { case 'C' : /* SIGCHLD */ zlog(ZLOG_DEBUG, "received SIGCHLD"); fpm_children_bury(); break; ...... case '2' : /* SIGUSR2 */ zlog(ZLOG_DEBUG, "received SIGUSR2"); zlog(ZLOG_NOTICE, "Reloading in progress ..."); fpm_pctl(FPM_PCTL_STATE_RELOADING, FPM_PCTL_ACTION_SET); break; } if (fpm_globals.is_child) { break; } } while (1); return; }
由下面的fpm_pctl()
代码可知,对于FPM_PCTL_ACTION_SET
操做只有当fpm状态fpm_state
为正常时(FPM_PCTL_STATE_NORMAL
),重启操做才能进行下去。
以后将重置已发送信号(fpm_signal_sent=0
),并设置fpm当前状态为FPM_PCTL_STATE_RELOADING
,而后调用fpm_pctl_action_next()
进行下一步操做。
void fpm_pctl(int new_state, int action) /* {{{ */ { switch (action) { case FPM_PCTL_ACTION_SET : if (fpm_state == new_state) { /* already in progress - just ignore duplicate signal */ return; } switch (fpm_state) { /* check which states can be overridden */ case FPM_PCTL_STATE_NORMAL : /* 'normal' can be overridden by any other state */ break; case FPM_PCTL_STATE_RELOADING : /* 'reloading' can be overridden by 'finishing' */ if (new_state == FPM_PCTL_STATE_FINISHING) break; case FPM_PCTL_STATE_FINISHING : /* 'reloading' and 'finishing' can be overridden by 'terminating' */ if (new_state == FPM_PCTL_STATE_TERMINATING) break; case FPM_PCTL_STATE_TERMINATING : /* nothing can override 'terminating' state */ zlog(ZLOG_DEBUG, "not switching to '%s' state, because already in '%s' state", fpm_state_names[new_state], fpm_state_names[fpm_state]); return; } fpm_signal_sent = 0; fpm_state = new_state; zlog(ZLOG_DEBUG, "switching to '%s' state", fpm_state_names[fpm_state]); /* fall down */ case FPM_PCTL_ACTION_TIMEOUT : fpm_pctl_action_next(); break; case FPM_PCTL_ACTION_LAST_CHILD_EXITED : fpm_pctl_action_last(); break; } }
此阶段能够当作是三个升级信号的发送过程:
SIGQUIT
信号,worker进程收到后会进行优雅关闭,并设置一个超时时为process_control_timeout
的定时器事件,关于process_control_timeout
能够看我另一篇文章【PHP】配置文件中的超时时间解析,定时器超时后最终将调用fpm_pctl(FPM_PCTL_STATE_UNSPECIFIED, FPM_PCTL_ACTION_TIMEOUT);
,从action名称能够看出是要进行超时的操做。fpm_pctl()
源码可知,action FPM_PCTL_ACTION_TIMEOUT
仍然调用fpm_pctl_action_next()
,只不过此次SIGQUIT
信号会升级为SIGTERM
发送给worker进程,定时器超时时间变为1s。SIGTERM
会升级为终极信号SIGKILL
。SIGKILL
信号相比SIGTERM
是不可被捕获或者忽略的,它将强行终止worker进程。static void fpm_pctl_action_next() /* {{{ */ { int sig, timeout; if (!fpm_globals.running_children) { fpm_pctl_action_last(); } if (fpm_signal_sent == 0) { if (fpm_state == FPM_PCTL_STATE_TERMINATING) { sig = SIGTERM; } else { sig = SIGQUIT; } timeout = fpm_global_config.process_control_timeout; } else { if (fpm_signal_sent == SIGQUIT) { sig = SIGTERM; } else { sig = SIGKILL; } timeout = 1; } // 实际调用kill() fpm_pctl_kill_all(sig); fpm_signal_sent = sig; fpm_pctl_timeout_set(timeout); }
worker进程主要处理master发送过来的三个信号,即SIGQUIT
、SIGTERM
、SIGKILL
。
SIGQUIT
信号的回调事件是sig_soft_quit()
。它首先会关闭listening_socket
,而且将in_shutdown
置为1,这样accept()
系统调用将当即返回-1,worker进程再也不接收请求,开始结束进程的操做。static void sig_soft_quit(int signo) /* {{{ */ { int saved_errno = errno; /* closing fastcgi listening socket will force fcgi_accept() exit immediately */ close(fpm_globals.listening_socket); if (0 > socket(AF_UNIX, SOCK_STREAM, 0)) { zlog(ZLOG_WARNING, "failed to create a new socket"); } // 设置in_shutdown=1 fpm_php_soft_quit(); errno = saved_errno; } int fcgi_accept_request(fcgi_request *req) { while (1) { if (req->fd < 0) { while (1) { if (in_shutdown) { return -1; } ...... req->fd = accept(listen_socket, (struct sockaddr *)&sa, &len); ...... } } else { fcgi_close(req, 1, 1); } } }
SIGTERM
信号采用SIG_DFL
默认处理方式,即终止进程,能够被阻塞、捕获、忽略。SIGKILL
信号不能被捕获或者忽略,将强行终止worker进程。worker进程的状态发生变化时,被终止或者暂停,内核会向master进程发送一个异步通知,即SIGCHLD
信号,由信号处理函数fpm_got_signal()
可知将执行fpm_children_bury()
。
下面将fpm_children_bury()
的代码拆解到对应部分下。
在这里先介绍下waitpid()
是干吗的:
当子进程结束的时候,内核会为终止子进程保存必定量的信息,这些信息至少包括进程ID、该进程的的终止状态、以及该进程使用的CPU时间总量。一个已经终止、可是其父进程还没有对其进行善后处理(获取终止子进程的有关信息,释放它仍占用的资源)的进程会成为僵尸进程。僵尸进程的进程号会被一直占用着,可是系统所能使用的进程号是有限的,因此若是有大量的僵尸进程产生,将由于没有可用的进程号而致使系统不能产生新的进程。
wait()
或waitpid()
就可让父进程获取到这些信息,并被内核释放掉。
// 最外层循环 while ( (pid = waitpid(-1, &status, WNOHANG | WUNTRACED)) > 0) { ...... }
master进程经过waitpid()
获取到终止的worker进程的pid
和终止状态status
后,将对status
进行一些判断
WTERMSIG(status)
来获取时子进程终止的信号编号。request_slowlog_timeout
后,master进程的心跳检测模块会给worker进程发送SIGSTOP
信号,worker进程被暂停,状态发生变化,内核向master进程发送SIGCHLD
信号,以后就会执行到这里。最后将调用fpm_php_trace()
函数来打印致使请求slow的堆栈信息。if (WIFEXITED(status)) { snprintf(buf, sizeof(buf), "with code %d", WEXITSTATUS(status)); /* if it's been killed because of dynamic process management * don't restart it automaticaly */ if (child && child->idle_kill) { restart_child = 0; } // 调用fpm_php_trace() if (WEXITSTATUS(status) != FPM_EXIT_OK) { severity = ZLOG_WARNING; } } else if (WIFSIGNALED(status)) { const char *signame = fpm_signal_names[WTERMSIG(status)]; const char *have_core = WCOREDUMP(status) ? " - core dumped" : ""; if (signame == NULL) { signame = ""; } snprintf(buf, sizeof(buf), "on signal %d (%s%s)", WTERMSIG(status), signame, have_core); /* if it's been killed because of dynamic process management * don't restart it automaticaly */ if (child && child->idle_kill && WTERMSIG(status) == SIGQUIT) { restart_child = 0; } if (WTERMSIG(status) != SIGQUIT) { /* possible request loss */ severity = ZLOG_WARNING; } } else if (WIFSTOPPED(status)) { zlog(ZLOG_NOTICE, "child %d stopped for tracing", (int) pid); if (child && child->tracer) { child->tracer(child); } continue; }
child = fpm_child_find(pid); if (child) { struct fpm_worker_pool_s *wp = child->wp; struct timeval tv1, tv2; // 资源释放 fpm_child_unlink(child); fpm_scoreboard_proc_free(wp->scoreboard, child->scoreboard_i); fpm_clock_get(&tv1); timersub(&tv1, &child->started, &tv2); ...... // 关闭标准输出、标准错误 fpm_child_close(child, 1 /* in event_loop */); // 在后文中详解 fpm_pctl_child_exited(); ...... } else { zlog(ZLOG_ALERT, "oops, unknown child (%d) exited %s. Please open a bug report (https://bugs.php.net).", pid, buf); }
从fpm_pctl_child_exited()
源码可知,若是这是最后一个worker进程的终止,将调用fpm_pctl(FPM_PCTL_STATE_UNSPECIFIED, FPM_PCTL_ACTION_LAST_CHILD_EXITED);
。
int fpm_pctl_child_exited() /* {{{ */ { if (fpm_state == FPM_PCTL_STATE_NORMAL) { return 0; } if (!fpm_globals.running_children) { fpm_pctl(FPM_PCTL_STATE_UNSPECIFIED, FPM_PCTL_ACTION_LAST_CHILD_EXITED); } return 0; }
继续追踪源码会发现,在重启操做中最后会调用fpm_pctl_exec()
。
execvp()
函数将从新执行php-fpm
程序,当前进程的正文、数据、堆和栈段都将被替换掉。
static void fpm_pctl_exec() /* {{{ */ { fpm_cleanups_run(FPM_CLEANUP_PARENT_EXEC); execvp(saved_argv[0], saved_argv); // 正常状况不会走到这里 zlog(ZLOG_SYSERROR, "failed to reload: execvp() failed"); exit(FPM_EXIT_SOFTWARE); }
至此,PHP-FPM就完成了重启。
PHP打印了不少Debug日志,你们能够在php-fpm.conf中将log_level
选项设置为debug
来开启。下面是debug日志的例子,能够对照着理解下上文内容。
[16-Jul-2019 16:51:40.248439] DEBUG: pid 36507, fpm_got_signal(), line 110: received SIGUSR2 [16-Jul-2019 16:51:40.248711] NOTICE: pid 36507, fpm_got_signal(), line 111: Reloading in progress ... [16-Jul-2019 16:51:40.248909] DEBUG: pid 36507, fpm_pctl(), line 229: switching to 'reloading' state [16-Jul-2019 16:51:40.249112] DEBUG: pid 36507, fpm_pctl_kill_all(), line 157: [pool www] sending signal 3 SIGQUIT to child 36508 [16-Jul-2019 16:51:40.249360] DEBUG: pid 36507, fpm_pctl_kill_all(), line 166: 1 child(ren) still alive [16-Jul-2019 16:51:40.249624] DEBUG: pid 36507, fpm_event_loop(), line 417: event module triggered 1 events [16-Jul-2019 16:51:40.256626] DEBUG: pid 36507, fpm_got_signal(), line 74: received SIGCHLD [16-Jul-2019 16:51:40.256968] DEBUG: pid 36507, fpm_children_bury(), line 259: [pool www] child 36508 exited with code 0 after 16.412179 seconds from start [16-Jul-2019 16:51:40.257411] NOTICE: pid 36507, fpm_pctl_exec(), line 96: reloading: execvp("/usr/local/Cellar/php/7.3.7/sbin/php-fpm", {"/usr/local/Cellar/php/7.3.7/sbin/php-fpm", "--fpm-config=/usr/local/etc/php/7.3.7/php-fpm.conf", "--pid=/usr/local/Cellar/php/7.3.7/var/run/php-fpm.pid"}) [16-Jul-2019 16:51:40.319184] DEBUG: pid 36507, fpm_unix_init_main(), line 518: The calling process is waiting for the master process to ping via fd=4 [16-Jul-2019 16:51:40.321064] DEBUG: pid 36699, fpm_scoreboard_init_main(), line 38: got clock tick '100' [16-Jul-2019 16:51:40.321588] NOTICE: pid 36699, fpm_sockets_init_main(), line 417: using inherited socket fd=7, "127.0.0.1:9001" [16-Jul-2019 16:51:40.321588] NOTICE: pid 36699, fpm_sockets_init_main(), line 417: using inherited socket fd=7, "127.0.0.1:9001" [16-Jul-2019 16:51:40.321782] DEBUG: pid 36699, fpm_socket_af_inet_socket_by_addr(), line 290: Found address for 127.0.0.1, socket opened on 127.0.0.1 [16-Jul-2019 16:51:40.321969] DEBUG: pid 36699, fpm_event_init_main(), line 335: event module is kqueue and 1 fds have been reserved [16-Jul-2019 16:51:40.322374] NOTICE: pid 36699, fpm_init(), line 83: fpm is running, pid 36699 [16-Jul-2019 16:51:40.322505] DEBUG: pid 36699, main(), line 1858: Sending "1" (OK) to parent via fd=5 [16-Jul-2019 16:51:40.322648] DEBUG: pid 36507, fpm_unix_init_main(), line 537: I received a valid acknowledge from the master process, I can exit without error [16-Jul-2019 16:51:40.322977] DEBUG: pid 36699, fpm_children_make(), line 428: [pool www] child 36702 started [16-Jul-2019 16:51:40.323302] DEBUG: pid 36699, fpm_event_loop(), line 364: 1296 bytes have been reserved in SHM [16-Jul-2019 16:51:40.323498] NOTICE: pid 36699, fpm_event_loop(), line 365: ready to handle connections