[apue] 使用文件记录锁无法实现父子进程交互执行同步
父子进程间交互执行是指用一种同步原语,实现父进程和子进程在某一时刻只有一个进程执行,之后由另外一个进程执行,用一段代码举例如下:
SYNC_INIT(); int i=0, counter=0; pid_t pid = fork (); if (pid < 0) err_sys ("fork error"); else if (pid > 0) { // parent for (i=0; i<NLOOPS; i+=2) { counter = update ((long *)area); if (counter != i) err_quit ("parent: expected %d, got %d", i, counter); else printf ("parent increase to %d based %d\n", i+1, counter); SYNC_TELL(pid, 1); SYNC_WAIT(0); } printf ("parent exit\n"); } else { for (i=1; i<NLOOPS+1; i+=2) { SYNC_WAIT(1); counter = update ((long *)area); if (counter != i) err_quit ("child: expected %d, got %d", i, counter); else printf ("child increase to %d based %d\n", i+1, counter); SYNC_TELL(getppid (), 0); } printf ("child exit\n"); }
其中area是指向共享内存的一个地址,update用来增加area指向的内容(为long),在fork之后,父子进程交替更新此值。
它们使用了一些抽象的同步原语,例如SYNC_INIT用于初始化同步设施、SYNC_WAIT等待另外进程的信号、SYNC_TELL向另外进程发送信号。
下面是成功同步后的输出(假设NLOOPS为100):
create shared-memory 3801126 with size 4 ok attach shared-memory at 0xb7733000 parent increase to 1 based 0 child increase to 2 based 1 parent increase to 3 based 2 child increase to 4 based 3 parent increase to 5 based 4 child increase to 6 based 5 parent increase to 7 based 6 child increase to 8 based 7 parent increase to 9 based 8 child increase to 10 based 9 parent increase to 11 based 10 child increase to 12 based 11 parent increase to 13 based 12 child increase to 14 based 13 parent increase to 15 based 14 child increase to 16 based 15 parent increase to 17 based 16 child increase to 18 based 17 parent increase to 19 based 18 child increase to 20 based 19 parent increase to 21 based 20 child increase to 22 based 21 parent increase to 23 based 22 child increase to 24 based 23 parent increase to 25 based 24 child increase to 26 based 25 parent increase to 27 based 26 child increase to 28 based 27 parent increase to 29 based 28 child increase to 30 based 29 parent increase to 31 based 30 child increase to 32 based 31 parent increase to 33 based 32 child increase to 34 based 33 parent increase to 35 based 34 child increase to 36 based 35 parent increase to 37 based 36 child increase to 38 based 37 parent increase to 39 based 38 child increase to 40 based 39 parent increase to 41 based 40 child increase to 42 based 41 parent increase to 43 based 42 child increase to 44 based 43 parent increase to 45 based 44 child increase to 46 based 45 parent increase to 47 based 46 child increase to 48 based 47 parent increase to 49 based 48 child increase to 50 based 49 parent increase to 51 based 50 child increase to 52 based 51 parent increase to 53 based 52 child increase to 54 based 53 parent increase to 55 based 54 child increase to 56 based 55 parent increase to 57 based 56 child increase to 58 based 57 parent increase to 59 based 58 child increase to 60 based 59 parent increase to 61 based 60 child increase to 62 based 61 parent increase to 63 based 62 child increase to 64 based 63 parent increase to 65 based 64 child increase to 66 based 65 parent increase to 67 based 66 child increase to 68 based 67 parent increase to 69 based 68 child increase to 70 based 69 parent increase to 71 based 70 child increase to 72 based 71 parent increase to 73 based 72 child increase to 74 based 73 parent increase to 75 based 74 child increase to 76 based 75 parent increase to 77 based 76 child increase to 78 based 77 parent increase to 79 based 78 child increase to 80 based 79 parent increase to 81 based 80 child increase to 82 based 81 parent increase to 83 based 82 child increase to 84 based 83 parent increase to 85 based 84 child increase to 86 based 85 parent increase to 87 based 86 child increase to 88 based 87 parent increase to 89 based 88 child increase to 90 based 89 parent increase to 91 based 90 child increase to 92 based 91 parent increase to 93 based 92 child increase to 94 based 93 parent increase to 95 based 94 child increase to 96 based 95 parent increase to 97 based 96 child increase to 98 based 97 parent increase to 99 based 98 child increase to 100 based 99 child exit parent exit remove that shared-memory
这套同步原语可以有多种实现方案,简单如管道、xsi信号量,甚至直接使用信号。下面是一些例子:
1. 使用管道
#ifdef USE_PIPE_SYNC // pp is the pipe that parent notify(write) child wait(read) // pc is the pipe that child notify(write) parent wait(read) static int pp[2], pc[2]; void SYNC_INIT (void) { if (pipe (pp) < 0 || pipe(pc) < 0) err_sys ("pipe error"); } void SYNC_TELL (pid_t pid, int child) { // close unused read end to avoid poll receive events // note, we can NOT do it in SYNC_INIT, // as at that moment, we have not fork yet ! if (child) { close (pp[0]); close (pc[1]); pp[0] = pc[1] = -1; } else { close (pc[0]); close (pp[1]); pc[0] = pp[1] = -1; } if (write (child ? pp[1] : pc[1], child ? "p" : "c", 1) != 1) err_sys ("write error"); } void SYNC_WAIT (int child /* unused */) { int n = 0, m = 0; struct pollfd fds[2] = {{ 0 }}; // if fd==-1, just be a place taker ! //if (pp[0] != -1) { fds[n].fd = pp[0]; fds[n].events = POLLIN; n++; } //if (pc[0] != -1) { fds[n].fd = pc[0]; fds[n].events = POLLIN; n++; } int ret = poll (fds, n, -1); if (ret == -1) err_sys ("poll error"); else if (ret > 0) { char c = 0; //printf ("poll %d from %d\n", ret, n); for (m=0; m<n; ++m) { //printf ("poll fd %d event 0x%08x\n", fds[m].fd, fds[m].revents); if (fds[m].revents & POLLIN) { if (fds[m].fd == pp[0]) { if (read (pp[0], &c, 1) != 1) err_sys ("read parent pipe error"); if (c != 'p') err_quit ("wait parent pipe but got incorrect data %c", c); } else { if (read (pc[0], &c, 1) != 1) err_sys ("read child pipe error"); if (c != 'c') err_quit ("wait child pipe but got incorrect data %c", c); } } } } else printf ("poll return 0\n"); } #endif
管道的话,TELL时就是向管道写一个字节,WAIT的时候就是阻塞在对应的端读取一个字节。
注意这里WAIT没有直接使用child参数,而是使用poll同时检测两个读端,看哪个有数据就返回哪个。其实直接读对应的端更直接一些。
2.使用xsi信号量
#ifdef USE_SEM_SYNC union semun { int val; //<= value for SETVAL struct semid_ds *buf; // <= buffer for IPC_STAT & IPC_SET unsigned short int *array;// <= array for GETALL & SETALL struct seminfo *__buf; // <= buffer for IPC_INFO }; static int semid = -1; void SYNC_INIT () { int mode = 0666; // 0; int flag = IPC_CREAT; #ifdef USE_EXCL flag |= IPC_EXCL; #endif semid = semget (IPC_PRIVATE, 2, flag | mode); if (semid < 0) err_sys ("semget for SYNC failed"); printf ("create semaphore %d for SYNC ok\n", semid); union semun sem; //sem.val = 1; //int ret = semctl (semid, 0, SETVAL, sem); //if (ret < 0) // err_sys ("semctl to set val failed"); short arr[2] = { 0 }; sem.array = arr; int ret = semctl (semid, 0, SETALL, sem); if (ret < 0) err_sys ("semctl to set all val failed"); printf ("reset all semaphores ok\n"); } void SYNC_TELL (pid_t pid, int child) { struct sembuf sb; sb.sem_op = 1; sb.sem_num = child ? 1 : 0; sb.sem_flg = 0; // IPC_NOWAIT, SEM_UNDO int ret = semop (semid, &sb, 1); if (ret < 0) printf ("semop to release resource failed, ret %d, errno %d\n", ret, errno); else printf ("release %d resource %d\n", sb.sem_op, ret); } void SYNC_WAIT (int child) { struct sembuf sb; sb.sem_op = -1; sb.sem_num = child ? 1 : 0; sb.sem_flg = 0; // IPC_NOWAIT, SEM_UNDO int ret = semop (semid, &sb, 1); if (ret < 0) printf ("semop to require resource failed, ret %d, errno %d\n", ret, errno); else printf ("require %d resource %d\n", sb.sem_op, ret); } #endif
xsi信号量的话,在TELL时是向对应的信号量执行V操作,释放一个资源;在WAIT时是向对应的信号量执行P操作,申请一个资源,如果申请不到,就阻塞在那里。
3.使用信号
#ifdef USE_SIGNAL_SYNC static volatile sig_atomic_t sigflag; static sigset_t newmask, oldmask, zeromask; static void sig_usr (int signo) { sigflag = 1; printf ("SIGUSR1/2 called\n"); } void SYNC_INIT () { if (apue_signal (SIGUSR1, sig_usr) == SIG_ERR) err_sys ("signal (SIGUSR1) error"); if (apue_signal (SIGUSR2, sig_usr) == SIG_ERR) err_sys ("signal (SIGUSR2) error"); sigemptyset (&zeromask); sigemptyset (&newmask); sigaddset (&newmask, SIGUSR1); sigaddset (&newmask, SIGUSR2); if (sigprocmask (SIG_BLOCK, &newmask, &oldmask) < 0) err_sys ("SIG_BLOCK error"); } void SYNC_TELL (pid_t pid, int child) { kill (pid, child ? SIGUSR1 : SIGUSR2); } void SYNC_WAIT (int child /* unused */) { while (sigflag == 0) sigsuspend (&zeromask); sigflag = 0; if (sigprocmask (SIG_SETMASK, &oldmask, NULL) < 0) err_sys ("SIG_SETMASK error"); } #endif
直接使用signal的话,这里分别使用了SIGUSR1和SIGUSR2表示父子进程,TELL操作就是激发一个信号给对方;WAIT操作就是sigsuspend在某个特定信号上,直到有信号发生才返回。
注意TELL时需要指定发送信号的进程号,所以多了一个pid参数,这个参数在之前据说的两种方法中并没有使用。这也是signal不好的一点。
然后,apue 15章最后一道习题中,要求使用文件记录锁来实现上述交互执行时,发现这是不可能完成的任务!
假设我们以加锁文件或文件中一个字节来实现WAIT,使用解锁来实现TELL,那么会发现文件记录锁有以下缺点,导致它不能胜任这个工作:
1. 文件记录锁是基于文件+进程的,当fork后产生子进程时,之前加的锁自动释放;
2. 文件记录锁对于重复施加锁于一个文件或文件中某个特定字节时,它的表现就和之前没有加锁一样,直接成功返回,不会产生阻塞效果;
对于 问题1,直接的影响就是父进程加好锁之后fork,子进程启动后却没有任何初始锁,导致父子进程同步困难。
虽然这个可以通过在子进程中重新初始化来部分的解决,但是这种问题因为有进程竞争存在,问题不严密从而不完美的;
对于 问题2,就直接导致其中一个进程在它的任务循环中,TELL另外一个进程后,再WAIT本进程的同步原语时(内部通过加锁实现),
另一个进程即使没有解锁相应的文件或字节,WAIT也直接成功返回(因为本进程已经持有该锁),从而造成其中一个进程执行多次,另一个进程没有办法插进去执行的情况(虽然两个进程也不能同时执行)。
所以结论是,对于交互执行的同步场景,管道、semaphore、signal都适用,而file lock不适用。
本文来自博客园,作者:goodcitizen,转载请注明原文链接:https://www.cnblogs.com/goodcitizen/p/file_locks_can_not_do_synchronizations.html