Tcl Source Code

The following script crashes both the HEAD and the 8.4
branch:
  open /dev/null
when run like this (from BASH, but there are other
equivalent situations):
  tclsh bug.tcl <&-

The effect of the invokation is to make tclsh run
without stdin opened (the same also happens with stdout
and stderr.)  This is caused by leaving the
initialisation of the channel table until the call to
Tcl_RegisterChannel() as part of the closing stages of
the [open] implementation, which ensures that fd0 (or 1
or 2) is put in the table twice, once for stdin/out/err
and once for the newly created channel for [open], with
the second time being the crash.

The attached patch fixes this, but causes test
encoding-24.3 to fail, and will break any script that
depends on setting [encoding system] early so as to
alter the encodings of the std channels.

This problem was discovered by rmax

Logged In: YES 
user_id=72656

test unixInit-7.1 {closed standard channel} -body {
    set tclsh [interpreter]
    makeFile {puts [open /dev/null]} crash.tcl
    makeFile [subst -nocommands {
close stdin
exec $tclsh crash.tcl
    }] crashtest.tcl
    exec $tclsh crashtest.tcl
} -returnCodes 0

in 8.4.6 and 8.5a1.

Logged In: YES 
user_id=72656

we should have a test case for this.

Logged In: YES 
user_id=72656

The alternate patch ignores Windows ... but is that 100% 
valid?  I recently had to solve this by checking for the 
standard channels being NULL on Windows.  I can't repro this 
with a variation of bug.tcl though.

File Added - 63206: open-std-channels.patch

Logged In: YES 
user_id=124643

How about this patch, at least as a workaround to prevent
the crash until a better solution is found? Unlike Donal's
proposal no tests fail with this patch applied. There might
be better places for adding these lines, but
TclpInitPlatform was the most reasonable place I have found.

Logged In: YES 
user_id=124643

Any news here, Jeff?
Today another colleague got hit by this bug, because stdin
is closed on processes started from fvwm2 buttons or menus.

Logged In: YES 
user_id=72656

I want a much better understanding of the implications of this 
bug before placing it into 8.4.x

File Added - 55966: channelbug.patch