诡异的bug: tcsh陷入死循环

2011-05-26 12:05 zhenjing 阅读(3167) 评论(2) 编辑收藏举报

问题：项目开发中，碰到一个很奇怪的问题：当tcsh启动的子程序退出之后，tcsh本身无法退出，并占用大量CPU资源。

背景：应用程序在fork之后，用tcsh启动另一个子进程，执行特定任务。进程之间使用sockepair(pipe)进行进程间通讯；为简化编程，将子进程的socket fd重定向为stdin和stdout。

具体症状：

Strace tcsh程序：

fstat(250, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(251, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(252, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(253, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(254, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(255, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) write(17, "[1] 30274\n", 10) = -1 EPIPE (Broken pipe) --- SIGPIPE (Broken pipe) @ 0 (0) --- lseek(16, 0, SEEK_END) = -1 ESPIPE (Illegal seek) rt_sigprocmask(SIG_BLOCK, [], [], 8) = 0 rt_sigprocmask(SIG_SETMASK, [], [], 8) = 0 fstat(0, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(1, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(2, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(3, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(4, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(5, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(6, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(7, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(8, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(9, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(10, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(11, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(12, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(13, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(14, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(20, 0x7fbfffc240) = -1 EBADF (Bad file descriptor) fstat(21, 0x7fbfffc240) = -1 EBADF (Bad file descriptor)

tcsh不断打印上述系统调用，可判断tcsh陷入了死循环，无法退出。其中

write(17, "[1] 30274\n", 10) = -1 EPIPE (Broken pipe)

--- SIGPIPE (Broken pipe) @ 0 (0) ---

lseek(16, 0, SEEK_END) = -1 ESPIPE (Illegal seek)

初步这两个系统调用是引发死循环的关键。为什么write到17文件描述符。(如果有经验，估计在这一步就可知道问题的根源。)

Lsof tcsh进程：

COMMAND   PID    USER   FD   TYPE             DEVICE     SIZE      NODE NAME
csh     30246 tachyon  cwd    DIR              253,7     4096 108511249 /h/work/tcpproxy
csh     30246 tachyon  rtd    DIR              253,0    20480         2 /
csh     30246 tachyon  txt    REG              253,0   349696    163890 /bin/tcsh
csh     30246 tachyon  mem    REG              253,3 48524688   1951970 /usr/lib/locale/locale-archive
csh     30246 tachyon  mem    REG              253,3    44473   1757371 /usr/share/locale/en/LC_MESSAGES/tcsh
csh     30246 tachyon  mem    REG              253,3    21546   1425608 /usr/lib64/gconv/gconv-modules.cache
csh     30246 tachyon  mem    REG              253,3    10128   1425550 /usr/lib64/gconv/ISO8859-1.so
csh     30246 tachyon  mem    REG              253,0   105080    196842 /lib64/ld-2.3.4.so
csh     30246 tachyon  mem    REG              253,0  1493409    196843 /lib64/tls/libc-2.3.4.so
csh     30246 tachyon  mem    REG              253,0    15728    196865 /lib64/libtermcap.so.2.0.8
csh     30246 tachyon  mem    REG              253,0    30070    196861 /lib64/libcrypt-2.3.4.so
csh     30246 tachyon   16u  unix 0x0000010293993640           31854667 socket
csh     30246 tachyon   17u  unix 0x0000010293993640           31854667 socket
csh     30246 tachyon   18u   REG              253,7 25770181 108511260 /****.log
csh     30246 tachyon   19u  unix 0x0000010293993640           31854667 socket

tcsh没有0，1，2的文件描述符，只有3个指向同一个unix地址的socket地址，基本断定： 16为stdin，17为stdout，18为stderr，19未知。

问题就变成：tcsh不断尝试往stdout输入特定信息！[1] 30274这个信息，应该是tcsh所起子进程的PID。

这样就基本清楚了：tcsh收到子进程的状态信息([1] PID)，并将收到的信息输出。由于stdout已被重定向成unix socket，在输出时，发生SIGPIPE (Broken pipe)事件。 SIGPIPE的默认行为是终止程序，显然tcsh没终止，并不断尝试输出。为什么tcsh不终止？难道tcsh内部实现屏蔽了该信号？为了解决这个问题，有两个办法：看tcsh的源码；看是否自己的程序改变了tcsh的信号处理。

最终，从自己的程序入手，发现果然是自己的程序改变了tcsh的信号配置。

A child created via fork(2) inherits a copy of its parent's signal dispositions. During an execve(2), the dispositions of handled signals are reset to the default; the dispositions of ignored signals are left unchanged.

问题的结论：

1) 自己的程序忽略了SIGPIPE；

2) fork之后，重定向了stdin和stdout;

3) tcsh启动子程序，捕获到子进程的状态(后台启动返回PID信息)，并尝试输出到stdout(socket)，因socket已被关闭，写失败并导致SIGPIPE。因SIGPIPE已被忽略，导致tcsh不断重试，无法退出！

重现程序：

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>

#include <sys/types.h>
#include <sys/socket.h>
#include <signal.h>

void Usage(char* program){
    fprintf(stderr, "Usage: %s binary\n", program);
    exit(-1);
}

void sighandler(int sig){
    /* fprintf(stderr, "recieve SIGPIPE\n"); */
}

int main(int argc, char** argv)
{
    if(argc < 2){
        Usage(argv[0]);
    }
    char* binary = argv[1];

    /* signal(SIGPIPE, sighandler); */
    signal(SIGPIPE, SIG_IGN);

    /* Get a socketpair we'll use for stdin and stdout. */
    int sp[2];
    if (socketpair(AF_UNIX, SOCK_STREAM, PF_UNSPEC, sp) == -1) {
        perror("socketpair error");
        return -1;
    }
    pid_t pid = fork();
    if (pid == 0) {
        signal(SIGTERM, SIG_DFL); signal(SIGHUP, SIG_DFL);
        close(sp[0]);
        if (dup2(sp[1], 0) == -1 || dup2(0, 1) == -1) {
            char msg[128];
            snprintf(msg, sizeof msg, "dup2 failed in child process %d", getpid());
        }

        sleep(1);
        fprintf(stderr, "start child process\n");
        setsid();
        execvp(binary, &argv[1]);
        char msg[512];
        snprintf(msg, sizeof msg, "exec %s failed", binary);
        exit(-1);
    }else if(pid < 0){
        perror("fork");
        exit(-1);
    }

    printf("parent process\n");
    close(sp[1]);
    close(sp[0]);

    while(1){
        char* msg = "hello test\n";
        /* write(sp[0], msg, strlen(msg)); */
        sleep(1);
        printf("write successful\n");
    }

    return 0;
}

gcc -o pipe pipe_broke.c
./pipe /usr/bin/strace -e trace=write,lseek -o strace.log csh -c "sleep 10 & "

Note: bash has no this problem! csh -> tcsh

刷新页面返回顶部

zhenjing的博客以文会友