php-fpm各進程運行模式

        之前一直以爲master進程負責監聽請求，當請求到來後，分發給worker進程進行處理。

        最近發現好像並不是這麼回事。先看fpm_main.c文件中的main()方法，裏面有這麼一段：

if (0 > fpm_init(argc, argv, fpm_config ? fpm_config : CGIG(fpm_config), fpm_prefix, fpm_pid, test_conf, php_allow_to_run_as_root, force_daemon, force_stderr)) {

	if (fpm_globals.send_config_pipe[1]) {
		int writeval = 0;
		zlog(ZLOG_DEBUG, "Sending \"0\" (error) to parent via fd=%d", fpm_globals.send_config_pipe[1]);
		zend_quiet_write(fpm_globals.send_config_pipe[1], &writeval, sizeof(writeval));
		close(fpm_globals.send_config_pipe[1]);
	}
	return FPM_EXIT_CONFIG;
}

...

fcgi_fd = fpm_run(&max_requests);
	
...

request = fpm_init_request(fcgi_fd);

        其中fpm_init()方法主要負責解析php-fpm.conf文件，獲取進程相關參數、初始化配置等；之後是fpm_run()方法，代碼如下：

/*	children: return listening socket
	parent: never return */
int fpm_run(int *max_requests) /* {{{ */
{
	struct fpm_worker_pool_s *wp;

	/* create initial children in all pools */
	for (wp = fpm_worker_all_pools; wp; wp = wp->next) {
		int is_parent;

		is_parent = fpm_children_create_initial(wp);

		if (!is_parent) {
			goto run_child;
		}

		/* handle error */
		if (is_parent == 2) {
			fpm_pctl(FPM_PCTL_STATE_TERMINATING, FPM_PCTL_ACTION_SET);
			fpm_event_loop(1);
		}
	}

	/* run event loop forever */
	fpm_event_loop(0);

run_child: /* only workers reach this point */

	fpm_cleanups_run(FPM_CLEANUP_CHILD);

	*max_requests = fpm_globals.max_requests;
	return fpm_globals.listening_socket;
}

        先看註釋可以知道，該方法返回了子進程worker監聽的套接字，但主進程master並不返回。看方法體，php-fpm有多個pool，每個pool下有多個worker進程，最終是調用了fpm_children_create_initial()方法裏的fpm_children_make()方法，在這個方法裏執行fork()生成了worker進程。生成子進程後，父進程就調用fpm_event_loop()方法進入無限循環狀態，而子進程則返回監聽的套接字並開始監聽。

        實驗如下，先看服務器運行的php-fpm進程：

[root@localhost src]# ps -ef | grep php
root     18204     1  0 Jun07 ?        00:00:15 php-fpm: master process (/usr/local/etc/php-fpm.conf)
www      18205 18204  0 Jun07 ?        00:00:04 php-fpm: pool www
www      18206 18204  0 Jun07 ?        00:00:04 php-fpm: pool www

        以上可以看到master進程pid爲18204，worker進程有兩個，pid分別爲18205和18206。通過strace跟蹤master進程：

[root@localhost fpm]# strace -t -p 18204
strace: Process 18204 attached
15:14:21 epoll_wait(9, [], 1, 484)      = 0
15:14:22 getsockopt(8, SOL_TCP, TCP_INFO, "\n\0\0\0\0\0\0\0@B\17\0\0\0\0\0\30\2\0\0\0\0\0\0\0\0\0\0\200\0\0\0"..., [104]) = 0
15:14:22 epoll_wait(9, [], 1, 1000)     = 0
15:14:23 getsockopt(8, SOL_TCP, TCP_INFO, "\n\0\0\0\0\0\0\0@B\17\0\0\0\0\0\30\2\0\0\0\0\0\0\0\0\0\0\200\0\0\0"..., [104]) = 0
15:14:23 epoll_wait(9, [], 1, 1000)     = 0
15:14:24 getsockopt(8, SOL_TCP, TCP_INFO, "\n\0\0\0\0\0\0\0@B\17\0\0\0\0\0\30\2\0\0\0\0\0\0\0\0\0\0\200\0\0\0"..., [104]) = 0
15:14:24 epoll_wait(9, [], 1, 1000)     = 0
15:14:25 getsockopt(8, SOL_TCP, TCP_INFO, "\n\0\0\0\0\0\0\0@B\17\0\0\0\0\0\30\2\0\0\0\0\0\0\0\0\0\0\200\0\0\0"..., [104]) = 0

        可以發現master進程會不斷循環調用epoll_wait()和getsockopt()兩個方法，用來處理異步信號事件和定時器事件。假如這時kill掉一個worker進程：

kill -9 18205

        再看跟蹤的master進程會有如下變化：

15:14:39 epoll_wait(9, 0x2193460, 1, 1000) = -1 EINTR (Interrupted system call)
15:14:39 --- SIGCHLD {si_signo=SIGCHLD, si_code=CLD_KILLED, si_pid=18205, si_uid=1001, si_status=SIGKILL, si_utime=396, si_stime=41} ---
15:14:39 write(7, "C", 1)               = 1
15:14:39 rt_sigreturn({mask=[]})        = -1 EINTR (Interrupted system call)
15:14:39 epoll_wait(9, [{EPOLLIN, {u32=17865216, u64=17865216}}], 1, 657) = 1
15:14:39 read(5, "C", 1)                = 1
15:14:39 wait4(-1, [{WIFSIGNALED(s) && WTERMSIG(s) == SIGKILL}], WNOHANG|WSTOPPED, NULL) = 18205
15:14:39 write(3, "[13-Jun-2020 15:14:39] WARNING: "..., 123) = 123
15:14:39 clone(child_stack=0, flags=CLONE_CHILD_CLEARTID|CLONE_CHILD_SETTID|SIGCHLD, child_tidptr=0x7fd55147ba10) = 23221
15:14:39 write(3, "[13-Jun-2020 15:14:39] NOTICE: ["..., 62) = 62
15:14:39 wait4(-1, 0x7ffd8e55a8fc, WNOHANG|WSTOPPED, NULL) = 0
15:14:39 read(5, 0x7ffd8e55a9df, 1)     = -1 EAGAIN (Resource temporarily unavailable)

        在執行epoll_wait發生中斷，最終調用clone()方法重新生成一個worker進程，並返回新子進程pid爲23221。處理完後繼續不斷調用epoll_wait()和getsockopt()倆方法。這是查看php-fpm進程進行驗證：

[root@localhost src]# ps -ef | grep php
root     18204     1  0 Jun07 ?        00:00:15 php-fpm: master process (/usr/local/etc/php-fpm.conf)
www      18206 18204  0 Jun07 ?        00:00:04 php-fpm: pool www
www      23221 18204  0 15:14 ?        00:00:00 php-fpm: pool www

        接下來再跟蹤worker進程，看看在幹嘛：

[root@localhost fpm]# strace -t -p 18206
strace: Process 18206 attached
16:04:32 accept(0, 

        發現worker進程阻塞在accept()方法這了。這時我發起一個請求，會看到如下變化：

[root@localhost fpm]# strace -t -p 18206
strace: Process 18206 attached
16:04:32 accept(0, {sa_family=AF_INET, sin_port=htons(32934), sin_addr=inet_addr("127.0.0.1")}, [16]) = 3
16:07:02 poll([{fd=3, events=POLLIN}], 1, 5000) = 1 ([{fd=3, revents=POLLIN}])
16:07:02 times({tms_utime=0, tms_stime=0, tms_cutime=0, tms_cstime=0}) = 807001318
16:07:02 read(3, "\1\1\0\1\0\10\0\0", 8) = 8
16:07:02 read(3, "\0\1\0\0\0\0\0\0", 8) = 8
16:07:02 read(3, "\1\4\0\1\3w\1\0", 8)  = 8
16:07:02 read(3, "\17\30SCRIPT_FILENAME/home/www/blog/"..., 888) = 888
16:07:02 read(3, "\1\4\0\1\0\0\0\0", 8) = 8
16:07:02 lstat("/home/www/blog/index.php", {st_mode=S_IFREG|0755, st_size=382, ...}) = 0
16:07:02 lstat("/home/www/blog", {st_mode=S_IFDIR|0755, st_size=4096, ...}) = 0
16:07:02 lstat("/home/www", {st_mode=S_IFDIR|0755, st_size=4096, ...}) = 0
16:07:02 lstat("/home", {st_mode=S_IFDIR|0755, st_size=4096, ...}) = 0
16:07:02 stat("/home/www/blog/.user.ini", 0x7ffcfb49b160) = -1 ENOENT (No such file or directory)
16:07:02 rt_sigaction(SIGPROF, NULL, {SIG_DFL, [], 0}, 8) = 0
...
16:07:02 rt_sigprocmask(SIG_UNBLOCK, [PROF], NULL, 8) = 0
16:07:02 mmap(NULL, 65536, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7f3fcbea7000
16:07:02 open("/home/www/blog/index.php", O_RDONLY) = 5
16:07:02 fstat(5, {st_mode=S_IFREG|0755, st_size=382, ...}) = 0
16:07:02 fstat(5, {st_mode=S_IFREG|0755, st_size=382, ...}) = 0
16:07:02 fstat(5, {st_mode=S_IFREG|0755, st_size=382, ...}) = 0
16:07:02 mmap(NULL, 382, PROT_READ, MAP_SHARED, 5, 0) = 0x7f3fcbf35000
16:07:02 getcwd("/usr/local/php/bin", 4095) = 19
16:07:02 chdir("/home/www/blog")        = 0
16:07:02 munmap(0x7f3fcbf35000, 382)    = 0
16:07:02 close(5)                       = 0
16:07:02 chdir("/usr/local/php/bin")    = 0
16:07:02 times({tms_utime=0, tms_stime=0, tms_cutime=0, tms_cstime=0}) = 807001319
16:07:02 stat("/etc/localtime", {st_mode=S_IFREG|0644, st_size=528, ...}) = 0
16:07:02 write(4, "127.0.0.1 -  13/Jun/2020:16:07:0"..., 61) = 61
16:07:02 setitimer(ITIMER_PROF, {it_interval={0, 0}, it_value={0, 0}}, NULL) = 0
16:07:02 write(3, "\1\6\0\1\0N\2\0X-Powered-By: PHP/7.2.4\r"..., 104) = 104
16:07:02 shutdown(3, SHUT_WR)           = 0
16:07:02 recvfrom(3, "\1\5\0\1\0\0\0\0", 8, 0, NULL, NULL) = 8
16:07:02 recvfrom(3, "", 8, 0, NULL, NULL) = 0
16:07:02 close(3)                       = 0
16:07:02 setitimer(ITIMER_PROF, {it_interval={0, 0}, it_value={0, 0}}, NULL) = 0
16:07:02 accept(0, 

        由上可以看出，是worker進程監聽socket，當請求過來時，worker進程對其進行處理，處理完後又重新調用accept()方法陷入阻塞狀態。