Evaluation of commands goes completely haywire if a file containing
a break/continue/return command outside its "intended" scope is sourced
using a dot command inside its "intended" scope. The main symptom is
not exiting from the sourced file when supposed to, leading to evaluation
of commands that were not supposed to be evaluated. A secondary symptom
is that these extra commands are not evaluated correctly, as some of them
are skipped. Some examples are listed in the How-To-Repeat section.
According to the POSIX standard, this is how it should work:
dot:
The shell shall execute commands from the file in the current
environment.
break:
The break utility shall exit from the smallest enclosing for, while,
or until loop, [...]
continue:
The continue utility shall return to the top of the smallest
enclosing for, while, or until loop, [...]
return:
The return utility shall cause the shell to stop executing
the current function or dot script. If the shell is not currently
executing a function or dot script, the results are unspecified.
It is clear that return should return from a sourced file, which
it does not do. Whether break and continue should work from the sourced
file might be debatable. Because the dot command says "in the current
environment", I'd say yes. In any case, it should not fail in weird
ways like it does now!
The problems occur with return (a) and break/continue (b) because:
1) dotcmd() does not record the function nesting level prior to
sourcing the file nor does it touch the loopnest variable,
leading to either
2 a) returncmd() being unable to detect that it should not set
evalskip to SKIPFUNC but SKIPFILE, or
b) breakcmd() setting evalskip to SKIPCONT or SKIPBREAK,
leading to
3) cmdloop() not detecting that it should skip the rest of
the file, due to only checking for SKIPFILE.
The result is that cmdloop() keeps executing lines from the file
whilst evalskip is set, which is the main symptom. Because
evalskip is checked in multiple places in eval.c, the secondary
symptom appears.
>How-To-Repeat:
Run the following script:
printf "break\necho break1; echo break2" >break
printf "continue\necho continue1; echo continue2" >continue
printf "return\necho return1; echo return2" >return
while true; do . ./break; done
for i in 1 2; do . ./continue; done
func() {
. ./return
}
func
No output should be produced, but instead this is the result:
break1
continue1
continue1
return1
The main symptom is evident from the unexpected output and the secondary
one from the fact that there are no lines with '2' in them.
>Fix:
Here is patch to src/bin/sh to fix the above problems. It keeps
track of the function nesting level at the beginning of a dot command
to enable the return command to work properly.
I also changed the undefined-by-standard functionality of the return
command when it's not in a dot command or function from (indirectly)
exiting the shell to being silently ignored. This was done because
the previous way has at least one bug: the shell exits without asking
for confirmation when there are stopped jobs.
Because I read the standard to mean that break and continue should have
an effect outside the sourced file, that's how I implemented it. For what
it's worth, this also seems to be what bash does. Also laziness, because
this way required no changes to loopnesting tracking. If this is not
wanted, it might make sense to move the nesting tracking to the inputfile
stack.
The patch also does some clean-up to reduce the amount of global
variables by moving the dotcmd() and the find_dot_file() functions from
main.c to eval.c and making in_function() a proper function.
sh +nounset and `for X; do` iteration fails if parameter set empty
by applying and testing FreeBSD's patch of Oct 24 2009 for this; see
http://svnweb.freebsd.org/base/head/bin/sh/expand.c?r1=198453&r2=198454
Also created an ATF test in tests/bin/sh/t_expand.sh for this error and
corrected a space->tabs problem there as well.
XXX: someone should fix all the .Ev stuff because some of them are just
shell variables .Va and are not really exported to the environment. See
the FreeBSD man page.
this fix was taken from FreeBSD SVN rev 199953 (Jilles Tjoelker)
------------------------------------------------------------------------
r199953 | jilles | 2009-11-30 07:33:59 +0900 (Mon, 30 Nov 2009) | 16 lines
Fix some cases where file descriptors from redirections leak to programs.
- Redirecting fds that were not open before kept two copies of the
redirected file.
sh -c '{ :; } 7>/dev/null; fstat -p $$; true'
(both fd 7 and 10 remained open)
- File descriptors used to restore things after redirection were not
set close-on-exec, instead they were explicitly closed before executing
a program normally and before executing a shell procedure. The latter
must remain but the former is replaced by close-on-exec.
sh -c 'exec 7</; { exec fstat -p $$; } 7>/dev/null; true'
(fd 10 remained open)
The examples above are simpler than the testsuite because I do not want to
use fstat or procstat in the testsuite.
the TOOLDIR version of libnbcompat, associated include files,
and associated defs.mk file, instead of the version from the
.OBJDIR of src/tools/compat. This should fix PR 47188.
- improve (create more consistency in) spelling;
- remove unnecessary (and in part ignored) macros, as well as an
unnecessary argument to `.Bl' (fixes mandoc(1) warnings);
- improve wording;
- bump date.
Patch from Bug Hunting.
if the block has moved, arrange so that trailing newlines are never placed in the string
in the first place, by accumulating them and adding them only after we've encountered a
non-newline character. This allows also for more efficient appending since we know how much
we need beforehand. From FreeBSD.
christos