Command substitutions containing a single simple command and here-document
expansion are performed in a subshell environment, but may not fork. Any
modified state of the shell environment should be restored afterward.
The state that OPTIND=1 had been done was not saved and restored here.
Note that the other parts of shellparam need not be saved and restored,
since they are not modified in these situations (a fork is done before such
modifications).
With the new expand.c code, the intermediate representation passed to the
pathname generation code only contains CTLESC, not CTLQUOTEMARK.
CTLQUOTEMARK now only occurs in the text of NARG nodes (output of the
parser).
This avoids the need to add and remove CTLESC bytes if pathname generation
will not be performed (set -f).
Side effect: the order of operations is slightly different: pathname
generation in ${$+* $(CMD)} will not see filesystem changes from CMD.
This simplifies the code and should be faster in some cases.
Side effect: the order of operations is different so that the value of IFS
used when IFS is modified during expansion (${IFS:=...}, ${IFS=...} or
$((...IFS=...))) may be different. Note that this order is highly unportable
between shells.
netbsd-tests.test.mk (r289151)
- Eliminate explicit OBJTOP/SRCTOP setting
- Convert all ad hoc NetBSD test integration over to netbsd-tests.test.mk
- Remove unnecessary TESTSDIR setting
- Use SRCTOP where possible for clarity
MFC after: 2 weeks
Sponsored by: EMC / Isilon Storage Divison
This simplifies the code (e.g. allowing use of qsort(3) instead of a
hand-rolled mergesort) and should have better cache properties.
The waste of unused args arrays after resizes is approximately the same as
the savings from getting rid of the next pointers.
At the same time, remove a piece of global state and move some duplicated
code into a function.
Shell syntax is too complicated to detect command substitution and unquoted
operators reliably without implementing much of sh's parser. Therefore, have
sh do this detection.
While changing sh's support anyway, also read input from a pipe instead of
arguments to avoid {ARG_MAX} limits and improve privacy, and output count
and length using 16 instead of 8 digits.
The basic concept is:
execl("/bin/sh", "sh", "-c", "freebsd_wordexp ${1:+\"$1\"} -f "$2",
"", flags & WRDE_NOCMD ? "-p" : "", <pipe with words>);
The WRDE_BADCHAR error is still implemented in libc. POSIX requires us to
fail strings containing unquoted braces with code WRDE_BADCHAR. Since this
is normally not a syntax error in sh, there is still a need for checking
code in libc, we_check().
The new we_check() is an optimistic check that all the characters
<newline> | & ; < > ( ) { }
are quoted. To avoid duplicating too much sh logic, such characters are
permitted when quoting characters are seen, even if the quoting characters
may themselves be quoted. This code reports all WRDE_BADCHAR errors; bad
characters that get past it and are a syntax error in sh return WRDE_SYNTAX.
Although many implementations of WRDE_NOCMD erroneously allow some command
substitutions (and ours even documented this), there appears to be code that
relies on its security (codesearch.debian.net shows quite a few uses).
Passing untrusted data to wordexp() still exposes a denial of service
possibility and a fairly large attack surface.
Reviewed by: wblock (man page only)
MFC after: 2 weeks
Relnotes: yes
Security: fixes command execution with wordexp(untrusted, WRDE_NOCMD)
POSIX requires this to prevent entering function definitions in history but
this implementation does nothing except retain the option's value. In ksh88,
function definitions were usually entered in the history file, even when
they came from ~/.profile and the $ENV file, to allow displaying their
definitions.
This is also the first option that does not have a letter.
The initial check for a matching ] was incorrect if a ] may be consumed by a
[:class:]. The subsequent loop assumed that there must be a ].
Remove the initial check and make the loop cope with a missing ].
Found with afl-fuzz.
MFC after: 1 week
An invalid substitution like ${var@} does not cause a parse error but is
stored in the intermediate representation, to be written as part of the
error message. If there is a CTL* byte in the stored part, this confuses
some code such as the code to skip an unused alternative such as in
${var-alternative}.
To keep things simple, do not store CTL* bytes.
Found with afl-fuzz.
MFC after: 1 week
The negative value was not expected and generated the low 8 bits as a byte,
which may be an invalid character encoding.
The final shift in creating the negative value was undefined as well.
Make the temporary variable unsigned to fix this.
Fix shifts of possibly negative numbers found with ubsan and avoid signed
integer overflow when hashing an extremely long command name.
MFC after: 1 week
Off by default, build behaves normally.
WITH_META_MODE we get auto objdir creation, the ability to
start build from anywhere in the tree.
Still need to add real targets under targets/ to build packages.
Differential Revision: D2796
Reviewed by: brooks imp
interactive.
I added the interactive check in r208881 to be safe, but in actual use
(scripts in set -m mode) passing along SIGINT seems best.
Discussed with: bdrewery
The parser considered 'trap exit INT' to reset the default for both EXIT and
INT. This beahvior is not POSIX compliant. This was avoided if a value was
specified for 'exit', but then disallows exiting with the signal received. A
possible workaround is using ' exit'.
However POSIX does allow this type of behavior if the parameters are all
integers. Fix the handling for this and clarify its support in the manpage
since it is specifically allowed by POSIX.
Differential Revision: https://reviews.freebsd.org/D2325
Reviewed by: jilles
MFC after: 2 weeks
EXP_REDIR was supposed to generate pathnames in redirection if exactly one
file matches, as permitted but not required by POSIX in interactive mode. It
is unlikely this will be implemented.
No functional change is intended.
MFC after: 1 week
Commands like 'export -p', 'set' and 'trap', and tracing enabled via 'set
-x' generate output suitable as shell input by adding quotes as necessary.
If there are control characters other than newline or invalid UTF-8
sequences, use $'...' and \OOO to display them safely.
The resulting output is not parsable by a strict POSIX.1-2008 shell but sh
from FreeBSD 9.0 and newer and many other shells can parse it.
Per Austin Group issue #459, shifting zero positional parameters may or may
not be considered an operand error (which causes the shell to exit in most
cases).
EXP_REDIR was not being checked for while expanding positional parameters in
redirection, so CTL* bytes were not being prefixed where they should be.
MFC after: 1 week
POSIX does not permit to continuing a getopts loop with different
arguments. For parsing the positional parameters, we handle this case by
resetting the getopts state when the positional parameters are changed in
any way (and the getopts state is local to a function). However, in the
syntax getopts <optstring> <var> <arg...>, changes could lead to invalid
memory access.
In the syntax getopts <optstring> <var> <arg...>, store a copy of the
arguments and continue to use them until getopts is reset.
These paths have had to be adjusted to changes in the testsuite runner
several times, so modify the tests to remove the need for such adjustment.
A cp in functional_test.sh is now unneeded, but this matters little in
performance.
In C, shift distances equal to or larger than the number of bits in the
operand result in undefined behaviour. As part of eliminating undefined
behaviour in arithmetic, mask off the distance like Java and JavaScript
specify and C on x86 usually does.
Assumption: conversion from unsigned to signed retains the two's complement
bits.
Assumption: uintmax_t has no padding bits.
The new code uses a "test discovery mechanism" to determine
what tests are available for execution
The test shell can be specified via:
kyua test -v test_suites.FreeBSD.bin.sh.test_shell=/path/to/test/sh
Sponsored by: EMC / Isilon Storage Division
Approved by: jmmv (mentor)
Reviewed by: jilles (maintainer)
variants. This allows usable file system images (i.e. those with both a
shell and an editor) to be created with only one copy of the curses library.
Exp-run: antoine
PR: 189842
Discussed with: bapt
Sponsored by: DARPA, AFRL
Currently, there can be no more than INT_MAX positional parameters. Make
sure to treat all higher ones as unset to avoid incorrect results and
crashes.
On 64-bit systems, our atoi() takes the low 32 bits of the strtol() and
sign-extends them.
On 32-bit systems, the call to atoi() returned INT_MAX for too high values
and there is not enough address space for so many positional parameters, so
there was no issue.
Although it is probably unwise to use this, POSIX is clear that leading
zeroes are permitted in positional parameters (and do not indicate octal).
Such positional parameters are checked for being unset and/or null
correctly, but their value is incorrectly expanded.
The test locale1.0 depends on locale support; it is meaningless without a
working LC_MESSAGES.
I added an OptionalObsoleteFiles.inc entry.
PR: 181151
Submitted by: Garrett Cooper (original version)
MFC after: 1 week
Sponsored by: EMC / Isilon Storage Division
When getopts finds an invalid option or a missing option-argument, it should
not reset its state and should set OPTIND as normal. This is an old ash bug
that was fixed long ago in dash. Our behaviour now matches most other
shells.
Only store exit status for a process if that process has not terminated yet.
Test (slow):
exit 7 & p1=$!; until exit 8 & p2=$!; [ "$p1" = "$p2" ]; do wait "$p2";
done; sleep 0.1; wait %1; echo $?
should write "7".
When killing a %job started without job control, kill all processes in it.
As with process groups and zombies, if any process in the job can be killed
or has already terminated, the command is successful.
This also fixes occasional failures of the builtins/kill1.0 test.
Change {atf,plain,tap}.test.mk to be internal implementation details of
bsd.test.mk. Makefiles that build tests should now only include bsd.test.mk
and declaratively specify what they want to build, without worrying about
the internal implementation of the mk files.
The reason for this change is to permit building test programs of different
interfaces from a single directory, which is something I had a need for
while porting tests over from src/tools/regression/.
Additionally, this change makes it possible to perform some other requested
changes to bsd.test.mk in an easier manner. Coming soon.
If a job has terminated but is still known, silently do nothing when using
the kill builtin with the job specifier. Formerly, the shell called kill()
with the process group ID that might have been reused.
Redo expari() like evalvar(). This makes the logic more understandable and
avoids possible problems if arithmetic expansion occurs if CTLESC characters
are not generated (looking backwards for CTLARI is not generally possible in
that case but the old code tried anyway).
This adds an extra argstr() recursion.
If an alias's value ends with a space or tab, the next word is also
checked for aliases.
This is a POSIX feature. It is useful with utilities like command and
nohup (alias them to themselves followed by a space).
Add the space to avoid alias recursion when the alias is expanded, not when
it is added.
As a result, displaying an alias via command -v, command -V or type no
longer erroneously appends a space. Adjust the tests so they now require
this bug to be absent.
The SIGWINCH handler triggers breakage in libedit which is hard to fix; see
PR bin/169773.
Also, window size changes while a program is in foreground (and it rather
than sh will receive SIGWINCH) will now be picked up automatically.
Downside: it is now certain that a resize is only processed after pressing
<Enter>. If libedit is fixed, sh will most likely have to be changed also.
PR: bin/180146
Redo r260506 by using the new TEST_METADATA functionality of bsd.test.mk
to mark the sh(1) and test(1) tests as not supporting root. This is to
get rid of hand-crafted Kyuafiles for these very simple cases.
MFC after: 5 days
One of the tests for test(1) fails and some of the tests for sh(1) are
silently bypassed when running as root.
To fix these tests and ensure they all run, mark the test programs for
sh(1) and test(1) as requiring an unprivileged user. (This should and
will be the default in Kyua but isn't yet.)
MFC after: 1 week
preadbuffer() maintained a flag whether there was any non-whitespace
character. This flag is only useful when history is enabled (in that case,
lines containing only whitespace are not added to history). Instead, check
using strspn() when history is enabled.
There is an approximate 2% speedup when running
sh -c '. /etc/rc.subr; . /etc/defaults/rc.conf; source_rc_confs'
with hot cache.
This change is a proof of concept on how to easily integrate existing
tests from the tools/regression/ hierarchy into the /usr/tests/ test
suite and on how to adapt them to the new layout for src.
To achieve these goals, this change:
- Moves tests from tools/regression/bin/<tool>/ to bin/<tool>/tests/.
- Renames the previous regress.sh files to legacy_test.sh.
- Adds Makefiles to build and install the tests and all their supporting
data files into /usr/tests/bin/.
- Plugs the legacy_test test programs into the test suite using the new
TAP backend for Kyua (appearing in 0.8) so that the code of the test
programs does not have to change.
- Registers the new directories in the BSD.test.dist mtree file.
Reviewed by: freebsd-testing
Approved by: rpaulo (mentor)
Although <&0 does nothing, it is a redirection affecting standard input and
should therefore disable the </dev/null redirection implicit in a background
command.
If job control is not enabled, background jobs started with ... & ignore
SIGINT and SIGQUIT so that they are not affected by such signals that are
intended for the foreground job. However, this should not prevent
reassigning a different action for these signals (as if the shell invocation
inherited these signal actions from its parent).
Austin group issue #751
Example:
{ trap - INT; exec sleep 10; } & wait
A Ctrl+C should terminate the sleep command.
user. Kqueue now saves the ucred of the allocating thread, to
correctly decrement the counter on close.
Under some specific and not real-world use scenario for kqueue, it is
possible for the kqueues to consume memory proportional to the square
of the number of the filedescriptors available to the process. Limit
allows administrator to prevent the abuse.
This is kernel-mode side of the change, with the user-mode enabling
commit following.
Reported and tested by: pho
Discussed with: jmg
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks
The getpgrp() call is unnecessary: if there is no job control then the
result was not used at all and if there is job control then we are not a
subshell and our process group ID is equal to our process ID (rootpid).
Formerly, return always returned from a function if it was called from a
function, even if there was a closer dot script. This was for compatibility
with the Bourne shell which only allowed returning from functions.
Other modern shells and POSIX return from the function or the dot script,
whichever is closest.
Git 1.8.4's rebase --continue depends on the POSIX behaviour.
Reported by: Christoph Mallon, avg
The change in r238888 was incomplete. It was still possible for a trapped
signal to arrive before the shell went to sleep (sigsuspend()) because a
check was missing or because the signal arrived before in_waitcmd was set.
On SMP, this bug sometimes caused the builtins/wait4.0 test to take 1 second
to execute; it then might or might not fail. On UP, the test almost always
failed.
The erflag argument was only used by old-style (``) command substitutions.
We can remove it and handle the special case in the command substitution
code.
NEOF needs to be a non-null pointer distinct from valid union node pointers.
It is not dereferenced.
The new NEOF is much like SIG_ERR except that it is an object pointer
instead of a function pointer.
The variable tokpushback can now be static.
As per POSIX, a simple command must have at least one redirection,
assignment word or command word.
These occured in rare cases such as eval "f()" .
The extension of allowing no commands inside { }, if, while, for, etc.
remains.
POSIX does not require ++ and -- in arithmetic. It is probably more useful
to reject them than to treat ++x and --x as x silently.
Note that the behaviour of increment and decrement can be obtained via
(x+=1), ((x+=1)-1), (x-=1) and ((x-=1)+1).
PR: bin/176444
If a job is specified to 'wait', wait for it to complete. Formerly, in
interactive mode, the job was deleted if it stopped.
If no jobs are specified in interactive mode, 'wait' still waits for all jobs
to complete or stop.
In non-interactive mode, WUNTRACED is not passed to wait3() so stopped jobs
are not detected.
PR: bin/181435
Replace the RESET blocks with regular functions and a reset() function that
calls them all.
This code generation tool is unusual and does not appear to provide much
benefit. I do not think isolating the knowledge about which modules need to
be reset is worth an almost 500-line build tool and wider scope for
variables used by the reset functions.
Also, relying on reset functions is often wrong: the cleanup should be done
in exception handlers so that no stale state remains after 'command eval'
and the like.
These cleanup operations are not needed because they are already performed
after an optimized command substitution (whether there was an error or not).
Although using -i with -c does not seem very useful, it seems inappropriate
to read commands from the terminal in this case.
Side effect: if the -s -c extension is used and the -s option is turned off
using 'set +s' during the interactive part, the shell now exits after an
error or interrupt. Note that POSIX only specifies -s as option to sh, not
to set.
See also Austin Group issue #718.
This is required by POSIX, at least for pids that are not known child
processes.
Other problems with job specifications still cause wait to abort with
exit status 2.
PR: 176916
This is only part of the PR; the behaviour for unknown/invalid pids/jobs
remains unchanged (aborts the builtin with status 2).
PR: 176916
Submitted by: Vadim Goncharov
The linked list of stack marks may cause problems if the allocation stack is
used between an exception and a higher-level popstackmark(), as it may then
touch a stack mark that is local to a function which has returned.
Also, the adjustment compares to a pointer passed to realloc(), which is
undefined behaviour.
Instead of adjusting stack marks when reallocating stack blocks, ensure that
such an adjustment is never necessary by fixing a small piece of memory in
place at a stack mark. This also simplifies the code.
To avoid the problems reported in bin/175922, it remains necessary to call
setstackmark() after popstackmark() if the stack mark remains in use.
* If read -t times out, return status as if interrupted by SIGALRM
(formerly 1).
* If a trapped signal interrupts read, return status 128+sig (formerly 1).
* If [EINTR] occurs but there is no trap, retry the read (for example
because of a SIGWINCH in interactive mode).
* If a read error occurs, write an error message and return status 2.
As before, a variable assignment error returns 2 and discards the remaining
data read.
Use non-blocking I/O to write as much as the pipe will accept (often 64K,
but it can be as little as 4K), avoiding the need for the ugly PIPESIZE
constant. If PIPESIZE was set too high, a deadlock would occur.
It now passes WARNS=7 with clang on i386.
GCC 4.2.1 does not understand setjmp() properly so will always trigger
-Wuninitialized. I will not add the volatile keywords to suppress this.
In some other shells, things like $((a);(b)) are command substitutions.
Also, there are shells that have an extension ((ARITH)) that evaluates an
arithmetic expression and returns status 1 if the result is zero, 0
otherwise. This extension may lead to ambiguity with two subshells starting
in sequence.
If syntactically invalid job identifiers are to be taken as jobs that exited
with status 127, this should not apply to options, so that we can add
options later if need be.
This ensures 'return' in a trap returns the correct status to the caller.
If evalskip is not set or if it is overridden by a previous evalskip, keep
the old behaviour of restoring the exit status from before the trap.
If a stack mark is set while the current stack block is empty, the stack
block may move later on (because of realloc()) and the stack mark needs to
be updated. This updating does not happen after popstackmark() has been
called; therefore, call setstackmark() again if the stack mark is still
being used.
For some reason, this only affects a few users. I cannot reproduce it. The
situation seems quite rare as well because an empty stack block would
usually be freed (by popstackmark()) before execution reaches a
setstackmark() call.
PR: 175922
Tested by: KT Sin
Now it outputs fixed files, which use constants provided by the C standard
library to determine appropriate values for the target machine.
Before, mksyntax inspected the host machine which resulted in subtle
breakage if e.g. char is signed on the host and unsigned on the target such
as when cross-compiling on x86 for ARM.
Tested using -funsigned-char on amd64. Compiling build-tools without it and
sh itself with it causes various tests to fail without this change but not
with this change. With consistent -funsigned-char, tests pass with or
without this change.
The mksyntax program could be removed and syntax.c and syntax.h committed to
the repository.
Submitted by: Christoph Mallon
MFC after: 2 weeks
ISO/IEC 9899:1999 (E) 5.2.1p3 guarantees that the values of the characters
0123456789 are contiguous.
The generated syntax.c and syntax.h remain the same.
Submitted by: Christoph Mallon
Expand here documents at the same point other redirections are expanded but
use a non-fork subshell environment (like simple command substitutions) for
compatibility. Substitition errors result in an empty here document like
before.
As a result, a fork is avoided for short (<4K) expanded here documents.
Unexpanded here documents (with quoted end marker after <<) are not affected
by this change. They already only forked when >4K.
Side effects:
* Order of expansion is slightly different.
* Slow expansions are not executed in parallel with the redirected command.
* A non-fork subshell environment is subtly different from a forked process.
If a loop contained certain commands (such as redirected compound commands),
the temporary memory for the redirection was not freed between iterations of
the loop but only after the loop.
Put a stackmark in evaltree(), freeing memory whenever a node has been
evaluated. Some other stackmarks are then redundant; remove them.
Example:
while :; do { :; } </dev/null; done
Instead of rechecking relative paths for all hashed utilities after a cd,
track if any utility in cmdtable depends on a relative path in PATH.
If there is such a utility, cd clears the entire table.
As a result, the '*' in hash no longer happens.
Accessing sys_siglist directly requires rtld to copy it from libc to the sh
executable's BSS. Also, strsignal() will put in the signal number for
unknown signals (FreeBSD-specific) so we need not do that ourselves.
Unfortunately, there is no function for sys_signame.
If there is a write error on stdout, a message will be printed (to stderr)
and the exit status will be changed to 2 if it would have been 0 or 1.
PR: bin/158206
* The last character is not displayed.
* If the alias ends with itself (as a word), an infinite memory-eating loop
occurs.
If an alias is defined initially, a space is appended to avoid recursion but
this did not happen when an alias was later modified.
PR: bin/173418
Submitted by: Daniel F.
MFC after: 1 week
Although sufficient memory is available for a longer string in cmdname,
this is undefined behaviour anyway.
Side effect: for alignment reasons, an additional byte of memory is
allocated per hashed command.
This reduces code duplication and code size.
/usr/bin/printf is not affected.
Side effect: different error messages when certain builtins are passed
invalid options.
When waiting for child processes using "wait" or if "set -T" is in effect, a
signal interrupts the wait. Make sure there is no window where the signal
handler may be invoked (setting a flag) just before going to sleep.
There is a similar race condition in the shell language, but scripts can
avoid it by exiting from the trap handler or enforcing synchronization using
a fifo.
If SIGCHLD is not trapped, a signal handler must be installed for it. Only
install this handler for the duration of the wait to avoid triggering
unexpected [EINTR] errors elsewhere.
Note that for some reason only SIGINT and SIGQUIT interrupt a "wait"
command. This remains the case.
Examples:
export x=~
now expands the tilde
local y=$1
is now safe, even if $1 contains IFS characters or metacharacters.
For a word to "look like an assignment", it must start with a name followed
by an equals sign, none of which may be quoted.
The special treatment applies when the first word (potentially after
"command") is "export", "readonly" or "local". There may be quoting
characters but no expansions. If "local" is overridden with a function there
is no special treatment ("export" and "readonly" cannot be overridden with a
function).
If things like
local arr=(1 2 3)
are ever allowed in the future, they cannot call a "local" function. This
would either be a run-time error or it would call the builtin.
This matches Austin Group bug #351, planned for the next issue of POSIX.1.
PR: bin/166771
On recent versions of NetBSD's libedit, el_gets
now sets el_len to -1 on error so we can
distinguish between a NULL string and an error.
This fixes sh from exiting with newer versions
of libedit now allowing EINTR to return.
Obtained from: NetBSD
Reviewed by: jilles
MFC after: 3 weeks
1. Avoid a cd back into ${.CURDIR} to run mkbuiltins when we know make
will first cd into ${.OBJDIR}. Keep the cwd to what make sets it to.
2. Don't tell mkbuiltins where to write to (= ${.OBJDIR}), but where to
get sources from (= ${.CURDIR}). This to compensate for point 1.
This fixes a problem with bmake's mk files that optimize ${.OBJDIR} to
expand to "." after changing cwd, not taking into account that the
target is pretty much undoing that and not getting the full path to the
object tree anymore.
quotation. Also make sure we have the same amount of columns in each row as
the number of columns we specify in the head arguments.
Reviewed by: brueffer
This uses vfork() for simple commands and command substitutions containing a
single simple command, invoking an external program under certain conditions
(no redirections or variable assignments, non-interactive shell, no job
control). These restrictions limit the amount of code executed in a vforked
child.
There is a large speedup (for example 35%) in microbenchmarks. The
difference in buildkernel is smaller (for example 0.5%) but still
statistically significant. See
http://lists.freebsd.org/pipermail/freebsd-hackers/2012-January/037581.html
for some numbers.
The use of vfork() can be disabled by setting a variable named
SH_DISABLE_VFORK.
In the first command of a 'for', $? should be the exit status of the last
pipeline (command substitution in the word list or command before 'for'),
not always 0.
Before this fix, only the first statement of the trap was executed if
evalskip was set. This is for example the case when:
o "-e" is set for this shell
o a trap is set on EXIT
o a function returns 1 and causes the script to abort
Reviewed by: jilles
MFC after: 2 weeks
Also, rework evalcase() to not evaluate any tree. Instead, return the
NCLISTFALLTHRU node and handle it in evaltree().
Fixed bugs:
* If a ;& list with non-zero exit status is followed by an empty ;; or final
list, the exit status of the case command should be equal to the exit
status of the ;& list, not 0.
* An empty ;& case should not reset $?.
* If no pattern is matched, POSIX says the exit status shall be 0 (even if
there are command substitutions).
* If a pattern is matched and there are no command substitutions, the first
command should see the $? from before the case command, not always 0.
The errno message display added in r222292 did not take attempting to
cd to a non-directory or something that cannot be stat()ed into account.
PR: bin/164070
MFC after: 10 days
This improves performance for globs where a slash or another component
follows a component with metacharacters by eliminating unnecessary attempts
to open directories that are not.
This is not necessary: errors are already caught in evalbackcmd() and
forcelocal handles changes to variables.
Note that this depends on r223024.
MFC after: 4 weeks
