#4311 Error in exec ends stdin

obsolete: 8.4.19
closed-invalid
5
2009-03-01
2009-02-28
No

Tcl version: 8.4+
OS: Linux Red Hat 4, 5
Maybe, it is a Linux bug, but it appeared in Tcl's stdin handling. I got the problem in C, but I created a Tcl script that fails the same way.
If you execute a certain command though system()/exec which fails, stdin stream does not provide any more input except already buffered in C code. My Tcl script is more that 8K long because I initially got the problem on the very next byte after 8K, but attached test fails a bit earlier.
Important: file must be redirected to stdin:
> tclsh <xxx.tcl
Read(0) just returns 0 as if it is the end of pipe/socket. If you specify file name as an argument, the script works.
I cannot classify the problem. Just weird. Stdin stays open, /proc/pid/fd shows it, but 0 handle is damaged. Fork/clone is doing something bad to to the parent's stdin.
Any help is appreciated. I cannot depend on any developer who can call system() and break stdin. Yes, I know that such redirection seem to make no sense, but it does; it is just a simplified version.

Discussion

  • Yevgen Ryazanov

    Yevgen Ryazanov - 2009-02-28

    Script that does not reach the end

     
  • Alexandre Ferrieux

    Please attach an strace:

    strace -f -tt -o tra tclsh < xxx.tcl

     
  • Donal K. Fellows

    • labels: --> 50. Embedding Support
     
  • Donal K. Fellows

    • milestone: --> obsolete: 8.4.19
    • assigned_to: nobody --> ferrieux
     
  • Yevgen Ryazanov

    Yevgen Ryazanov - 2009-02-28

    strace -f -tt -o tra tclsh < xxx.tcl

     
  • Yevgen Ryazanov

    Yevgen Ryazanov - 2009-02-28

    File Added: tra

     
  • Yevgen Ryazanov

    Yevgen Ryazanov - 2009-02-28

    File Added: tra

     
  • Alexandre Ferrieux

    OK got it. This is not a Tcl nor Linux problem, just one specific behavior of csh.
    Details from the strace:

    12200 09:32:12.664493 execve("/bin/csh", ["csh"..., "-fc"..., "echo hello >fff"...], [/* 69 vars */] <unfinished ...>
    12200 09:32:12.675081 dup2(0, 16) = 16
    12200 09:32:12.747895 open("fff", O_WRONLY|O_CREAT|O_TRUNC, 0666) = -1 EACCES (Permission denied)
    12200 09:32:12.748463 write(18, "fff: Permission denied.\n", 24) = 24
    12200 09:32:12.748541 lseek(16, 0, SEEK_END) = 8660

    Explanation:

    (1) The dup2 is part of the redirections for the "echo hello" command. It just inherits desc 0 because there is no redirection on stdin.
    (2) Csh tries to honour the > redirection, and gets an (expected) error. It barfs to stderr (which is okay).
    (3) *but* it also decides to skip ahead to the end of the local subcommand's stdin (lseek(16,...))
    (4) All this hooks back to Tcl's stdin since the [exec csh] has no stdin redirection either.
    (5) Hence on the next attempt to read, the Tcl IO subsystem gets an EOF. QED.

    I have no idea why Csh does (3). At the extreme it could vaguely make sense for an interactive Csh to do that on a terminal, so that it skips to new typed command when an error has occurred in a buffered block of commands... But here the Csh is not interactive. From my seat it is a fully qualified Bug.

    A simple workaround is to [exec csh ... < /dev/null].

    I'll thus close this artifact as Invalid, since Tcl cannot be held responsible for Csh's misbehavior.
    My personal take on Csh is: only use it interactively if ever. That's not a programming language. That's an engineer's nightmare.
    *Always* use /bin/sh when you need to resort to an OS shell.

     
  • Alexandre Ferrieux

    • status: open --> closed-invalid
     
  • Yevgen Ryazanov

    Yevgen Ryazanov - 2009-03-01

    Thanks!!!

     

Get latest updates about Open Source Projects, Conferences and News.

Sign up for the SourceForge newsletter:





No, thanks