Hello Guus, hello Ziga I found the problem.
Guus Sliepen schrieb am Tue 27. Mar, 12:46 (+0200):
> On Thu, Mar 22, 2007 at 05:23:36PM +0100, Jörg Sommer wrote:
>
> > > The patch didn't work. I then reinstalled the original script, but
> > > removed all redirections to /dev/null. It still complains about not
> > > being able to open /dev/null.
> >
> > Oh, can you add a set -x before the PATH assignment and than look, if you
> > can see between which lines the message is printed. BTW: You can stop the
> > output with ctrl+s.
>
> I see the following:
>
> [...]
> + [ ! -f /proc/stat ]
> + sleep 0.2
> /sbin/bootchartd: 1: Can't open /dev/null
> /sbin/bootchartd: 1: Can't open /dev/null
> /sbin/bootchartd: 1: + [ -n yes ]
> /sbin/bootchartd: 1: Can't open /dev/null
> Can't open /dev/null
The lines above belong to these lines:
while [ ! -f /proc/stat ]; do sleep $SAMPLE_PERIOD; done
sleep $SAMPLE_PERIOD
# Run loggers in background
log_output "cat /proc/stat" proc_stat.log &
# /proc/diskstats is available in 2.6 kernels
[ -f /proc/diskstats ] && log_output "cat /proc/diskstats" proc_diskstats.log
&
log_output "cat /proc/*/stat" proc_ps.log &
This is a short form of the problamatical code:
#!/bin/dash
foo()
{
echo $!
/bin/ls -l /proc/$$/fd
/bin/sleep 2
}
foo &
foo
If you use strace, you can find these lines:
[pid 17961] clone(Process 17964 attached
child_stack=0, flags=CLONE_CHILD_CLEARTID|CLONE_CHILD_SETTID|SIGCHLD,
child_tidptr=0x3001fd68) = 17964
[pid 17961] wait4(-1, Process 17961 suspended
<unfinished ...>
[pid 17963] close(10) = 0
[pid 17963] rt_sigaction(SIGINT, {SIG_IGN}, {0x100115b0, ~[KILL STOP RTMIN
RT_1], 0}, 8) = 0
[pid 17963] rt_sigaction(SIGQUIT, {SIG_IGN}, {SIG_DFL}, 8) = 0
[pid 17963] close(0) = 0
[pid 17963] open("/dev/null", O_RDONLY) = 0
[pid 17963] write(1, "\n", 1
) = 1
And there you see, dash tries to open /dev/null. The corresponding code
is the following: (dash-0.5.3/src/jobs.c)
/*
* Fork off a subshell. If we are doing job control, give the subshell its
* own process group. Jp is a job structure that the job is to be added to.
* N is the command that will be evaluated by the child. Both jp and n may
* be NULL. The mode parameter can be one of the following:
* FORK_FG - Fork off a foreground process.
* FORK_BG - Fork off a background process.
* FORK_NOJOB - Like FORK_FG, but don't give the process its own
* process group even if job control is on.
*
* When job control is turned off, background processes have their standard
* input redirected to /dev/null (except for the second and later processes
* in a pipeline).
*
* Called with interrupts off.
*/
STATIC inline void
forkchild(struct job *jp, union node *n, int mode)
{
int oldlvl;
TRACE(("Child shell %d\n", getpid()));
oldlvl = shlvl;
shlvl++;
closescript();
clear_traps();
#if JOBS
/* do job control only in root shell */
jobctl = 0;
if (mode != FORK_NOJOB && jp->jobctl && !oldlvl) {
pid_t pgrp;
if (jp->nprocs == 0)
pgrp = getpid();
else
pgrp = jp->ps[0].pid;
/* This can fail because we are doing it in the parent also */
(void)setpgid(0, pgrp);
if (mode == FORK_FG)
xtcsetpgrp(ttyfd, pgrp);
setsignal(SIGTSTP);
setsignal(SIGTTOU);
} else
#endif
if (mode == FORK_BG) {
ignoresig(SIGINT);
ignoresig(SIGQUIT);
if (jp->nprocs == 0) {
close(0);
if (open(_PATH_DEVNULL, O_RDONLY) != 0)
sh_error("Can't open %s", _PATH_DEVNULL);
}
}
Three lines above you can see the problamatical line of code. The path is
hard coded and there's (no) escape. The only way I found out is job
control.
#!/bin/dash
exec 3>&1 <>/dev/null >&0
set -m
foo()
{
echo $!
/bin/ls -l /proc/$$/fd
/bin/sleep 2
}
echo hello >&3
foo &
foo
echo bye >&3
For job control bootchart needs a heavy rewrite, but I also think it can
gain something. The other way is to try to redirect the shell error to
/dev/null.
What do you think Ziga?
Bye, Jörg.
--
“Perl—the only language that looks the same
before and after RSA encryption.” (Keith Bostic)
pgpzB7eE2nZ3Y.pgp
Description: PGP signature
