| Commit message (Collapse) | Author |
|---|
|
|
|
Previously, the prompt would be expanded every time editline called the
getprompt callback. I think the code may have been written assuming that
editline only calls getprompt once per prompt, but it may actually call
it many times, for instance every time you type backspace. This results
not only in slower editing from expanding complex prompts repeatedly, it
also consumes more and more stack memory each time getprompt is called.
This can be seen by setting PS1 to some command substitution, typing
many characters at the prompt, then holding backspace and observing
memory usage. Thankfully all this stack memory is freed between prompts
by the stackmark calls around el_gets.
This change causes prompt expansion to always happen in the setprompt
call, as it would when editline is disabled, and a cached copy of the
prompt is saved for getprompt to return every time editline calls it.
Since getprompt is no longer doing expansion, the stackmark calls
surrounding el_gets can be removed.
|
|
Check if the FreeBSD (and by extension macOS) _el_fn_sh_complete version
of the function exists, which does shell escaping of completed filenames
before _el_fn_complete learned to.
Tab is bound after calling el_set() and el_source() since the default
bindings set tab to something else, and it should always be filename
completion in the shell.
|
|
|
|
|
|
On Sat, Jun 19, 2021 at 02:44:46PM +0200, Denys Vlasenko wrote:
>
> CTLVAR and CTLBACKQ are not properly handled if encountered
> inside {$#...}. Testcase:
>
> dash -c "`printf 'echo ${#1\x82}'`" 00 111 222
>
> It should execute "echo ${#1 <byte 0x82> }" and thus print "3"
> (the length of $1, which is "111").
>
> Instead, it segfaults.
>
> (Ideally, it should fail since "1 <byte 0x82>" is not a valid
> variable name, but currently dash accepts e.g. "${#1abc}"
> as if it is "${#1}bc". A separate, less serious bug...).
In fact these two bugs are one and the same. This patch fixes
both by detecting the invalid substitution and not emitting it
into the node tree.
Incidentally this reveals a bug in how we parse ${#10} that got
introduced recently, which is also fixed here.
Reported-by: Denys Vlasenko <vda.linux@googlemail.com>
Fixes: 7710a926b321 ("parser: Only accept single-digit parameter...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
|
|
Patrick Brünn <P.Bruenn@beckhoff.com> wrote:
>
> Since we are migrating to Debian bullseye, we discovered a new behavior
> with our scripts, which look like this:
>>#!/bin/sh
>>cleanup() {
>> set +e^M
>> rmdir ""
>>}
>>set -eu
>>trap 'cleanup' EXIT INT TERM
>>echo 'Hello world!'
>
> With old dash v0.5.10.2 this script would return 0 as we expected it.
> But since commit 62cf6955f8abe875752d7163f6f3adbc7e49ebae it returns
> the last exit code of our cleanup function.
> Reverting that commit gives a merge conflict, but it seems to fix _our_
> problem. As that topic appears too complex to us I want to ask the
> experts here:
>
> Is this change in behavior intended, by dash?
>
> Our workaround at the moment would be:
>>trap 'cleanup || true' EXIT INT TERM
Thanks for the report. This is actually a fairly old bug with
set -e that's just been exposed by the exit status change. What's
really happening is that cleanup itself is triggering a set -e
exit incorrectly because evaltree cached the value of eflag prior
to the function call.
This patch should fix the problem.
Reported-by: Patrick Brünn <P.Bruenn@beckhoff.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Tested-by: Patrick Brünn <P.Bruenn@beckhoff.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
|
|
This patch ensures that waitcmd never blocks unless there are
outstanding jobs. This could otherwise trigger a hang if children
were created prior to the shell coming into existence, or if
there are backgrounded children of other kinds (e.g., a here-
document).
Fixes: 6c691b3e5099 ("jobs: Only clear gotsigchld when waiting...")
Reported-by: Michael Biebl <biebl@debian.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
|
|
The test for open64 is separate from stat64 for macOS. However,
the newly introduced tests for readdir64/dirent64 should be grouped
with open64 instead of stat64 as otherwise they cause similar build
failures.
Reported-by: Martijn Dekker <martijn@inlv.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
|
|
We need to clear gotsigchld in waitproc because it is used as
a loop conditional for the waitcmd case. Without it waitcmd
may busy loop after a SIGCHLD.
This patch also changes gotsigchld into a volatile sig_atomic_t
to prevent compilers from optimising its accesses away.
Fixes: 6c691b3e5099 ("jobs: Only clear gotsigchld when waiting...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
|
|
When handling backslashes within an old-style command substitution,
we should not call pgetc_eatbnl because that would treat the next
backslash character as another escape character if it was then
followed by a new-line.
This patch fixes it by calling pgetc.
Reported-by: Matt Whitlock <dash@mattwhitlock.name>
Fixes: 6bbc71d84bea ("parser: use pgetc_eatbnl() in more places")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Signed-off-by: Martin Michlmayr <tbm@cyrius.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
On Sun, May 17, 2020 at 01:19:28PM +0100, Harald van Dijk wrote:
>
> This still does not restore the state completely. It does not clean up any
> pending heredocs. I see:
>
> $ PS1='$(<<EOF "'
> src/dash: 1: Syntax error: Unterminated quoted string
> $(<<EOF ":
> >
>
> That is, after entering the ':' command, the shell is still trying to read
> the heredoc from the prompt.
This patch saves and restores the heredoclist in expandstr.
It also removes a bunch of unnecessary volatiles as those variables
are only referenced in case of a longjmp other than one started by
a signal like SIGINT.
Reported-by: Harald van Dijk <harald@gigawatt.nl>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
There are some remaining stat/readdir calls in dash that may lead
to spurious EOVERFLOW errors on 32-bit platforms. This patch changes
them (as well as open(2)) to use the explicit large file API.
Reported-by: Tatsuki Sugiura <sugi@nemui.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
With -fno-common, which will be enabled by default in GCC 10, we see
this error:
ld: input.o:(.bss+0x0): multiple definition of `el';
histedit.o:(.bss+0x8): first defined here
To fix this, simply remove the definition as it is not needed.
Signed-off-by: Jeroen Roovers <jer@gentoo.org>
Signed-off-by: Mike Gilbert <floppym@gentoo.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The mktokens script fails when /tmp isn't writable (e.g., when building
in a sandbox with a different TMPDIR). Replace absolute references to
/tmp to relative references to TMPDIR. If TMPDIR is unset or null,
default to /tmp.
The mkbuiltins script was already hardened to work relative to TMPDIR,
also defaulting to /tmp.
v2 ensures that TMPDIR is quoted.
v3 adds an extra quotation that prevents extra pathname expansions.
Signed-off-by: Michael Greenberg <michael.greenberg@pomona.edu>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
This script should print OK:
alias a="case x in " b=x
a
b) echo BAD;; esac
alias BEGIN={ END=}
BEGIN
cat <<- EOF > /dev/null
$(:)
EOF
END
: <<- EOF &&
$(:)
EOF
BEGIN
echo OK
END
However, because the value of checkkwd is either zeroed when it
shouldn't, or isn't zeroed when it should, dash currently gets
it wrong in every case.
This patch fixes it by saving checkkwd and zeroing it where needed.
Suggested-by: Harald van Dijk <harald@gigawatt.nl>
Reported-by: Harald van Dijk <harald@gigawatt.nl>
Reported-by: Martijn Dekker <martijn@inlv.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
On Fri, Dec 13, 2019 at 02:51:34PM +0000, Simon Ser wrote:
> Just noticed another dash bug: when setting invalid PS1 values dash
> enters an infinite loop.
>
> For instance, setting PS1='$(' makes dash print many of these:
>
> dash: 1: Syntax error: end of file unexpected (expecting ")")
>
> It would be nice to fallback to the default PS1 value on error.
This patch fixes it by using the literal value of PS1 should an
error occur during expansion.
On Wed, Feb 26, 2020 at 09:12:04PM +0000, Ron Yorston wrote:
>
> There's another case that should be handled. PS1='`xxx(`' causes the
> shell to exit because the old-style backquote leaves an additional file
> on the stack.
Ron's change has been folded into this patch.
Reported-by: Simon Ser <contact@emersion.fr>
Reported-by: Ron Yorston <rmy@pobox.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Dash was incorrectly handling empty aliases. When attempting to use an
empty alias with nothing else, I'm (incorrectly) prompted for more
input:
```
$ alias empty=''
$ empty
>
```
Other shells (e.g., bash, yash) correctly handle the lone, empty alias as an
empty command:
```
$ alias empty=''
$ empty
$
```
The problem here is that we incorrectly enter the loop eating TNLs
in readtoken(). This patch fixes it by setting checkkwd correctly.
Reported-by: Michael Greenberg <michael.greenberg@pomona.edu>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
After we changed the resetting of gotsigchld so that it is only
done if jp is NULL, we can now get an infinite loop in waitproc
if gotsigchld is set but there is no outstanding child because
everything had been waited for previously without gotsigchld being
zeroed.
This patch fixes it by always zeroing gotsigchld as we did before.
The bug that the previous patch was trying to fix is now resolved
by switching the blocking mode to DOWAIT_NORMAL after the specified
job has been completed so that we really do wait for all outstanding
dead children.
Reported-by: Harald van Dijk <harald@gigawatt.nl>
Fixes: 6c691b3e5099 ("jobs: Only clear gotsigchld when waiting...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
To make it clearer what it is doing: nonblocking wait()
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The seconds part of the times(1) built-in is wrong as it does not
exclude the minutes part of the result. This patch fixes it.
This problem was first noted by Michael Greenberg who also sent
a similar patch.
Reported-by: Michael Greenberg <michael.greenberg@pomona.edu>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
When we enter a subshell we need to drop the saved redirections
as otherwise a subsequent unwindredir could produce incorrect
results.
This patch does this by simply clearing redirlist. While we
could actually free the memory underneath for subshells it isn't
really worth the trouble for now.
In order to ensure that this is done in every place where we enter
a subshell, this patch adds a new mkinit hook called forkreset.
The calls closescript, clear_traps and reset_handler are also added
to the forkreset hook.
This fixes a bug where the first two functions weren't called
if we enter a subshell without forking.
Reported-by: Harald van Dijk <harald@gigawatt.nl>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
On Thu, Apr 25, 2019 at 01:39:52AM +0000, Michael Orlitzky wrote:
> The POSIX spec says,
>
> The parameter name or symbol can be enclosed in braces, which are
> optional except for positional parameters with more than one digit or
> when parameter is a name and is followed by a character that could be
> interpreted as part of the name.
>
> However, dash seems to diverge from that behavior when we get to $10:
>
> $ cat test.sh
> echo $10
>
> $ dash ./test.sh one two three four five six seven eight nine ten
> ten
>
> $ bash ./test.sh one two three four five six seven eight nine ten
> one0
This patch should fix the problem.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
According to POSIX.1-2008 we should remove newlines only at the end of
the substitution. Newlines-only substitions causes dash to remove
newlines before beggining of the substitution. The following code:
cat <<END
1
$(echo "")
2
END
prints "1<newline>2" instead of expected "1<newline><newline>2".
This patch fixes trailing newlines processing in backquote expanding.
Signed-off-by: Nikolai Merinov <n.merinov@inango-systems.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
On Wed, Jul 25, 2018 at 12:38:27PM +0000, project-repo wrote:
> Hi,
> I am working on a project in which I use the honggfuzz fuzzer to fuzz open
> source software and I decided to fuzz dash. In doing so I discovered a
> NULL pointer dereference in src/redir.ch on line 305. Following is a
> backtrace as supplied by the address sanitizer:
>
> AddressSanitizer:DEADLYSIGNAL
> =================================================================
> ==39623==ERROR: AddressSanitizer: SEGV on unknown address 0x000000000010 (pc 0x0000005768ed bp 0x7ffc00273df0 sp 0x7ffc00273c60 T0)
> ==39623==The signal is caused by a READ memory access.
> ==39623==Hint: address points to the zero page.
> #0 0x5768ec in openhere /home/jfe/dash/src/redir.c:305:29
> #1 0x574d92 in openredirect /home/jfe/dash/src/redir.c:230:7
> #2 0x5737fe in redirect /home/jfe/dash/src/redir.c:121:11
> #3 0x576017 in redirectsafe /home/jfe/dash/src/redir.c:424:3
> #4 0x522326 in evalcommand /home/jfe/dash/src/eval.c:828:11
> #5 0x520010 in evaltree /home/jfe/dash/src/eval.c:288:12
> #6 0x5270da in evaltreenr /home/jfe/dash/src/eval.c:332:2
> #7 0x526f04 in evalbackcmd /home/jfe/dash/src/eval.c:640:3
> #8 0x539020 in expbackq /home/jfe/dash/src/expand.c:522:2
> #9 0x5332d7 in argstr /home/jfe/dash/src/expand.c:343:4
> #10 0x5322f7 in expandarg /home/jfe/dash/src/expand.c:196:2
> #11 0x528118 in fill_arglist /home/jfe/dash/src/eval.c:659:3
> #12 0x5213b6 in evalcommand /home/jfe/dash/src/eval.c:769:13
> #13 0x520010 in evaltree /home/jfe/dash/src/eval.c:288:12
> #14 0x554423 in cmdloop /home/jfe/dash/src/main.c:234:8
> #15 0x553bcc in main /home/jfe/dash/src/main.c:176:3
> #16 0x7f201c2b2a86 in __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x21a86)
> #17 0x41dfb9 in _start (/home/jfe/dash/src/dash+0x41dfb9)
>
> AddressSanitizer can not provide additional info.
> SUMMARY: AddressSanitizer: SEGV /home/jfe/dash/src/redir.c:305:29 in openhere
> ==39623==ABORTING
>
> This bug can be reproduced by running "dash < min" where min is þhe file
> attached. I was able to reproduce this bug with the current git version
> and the current debian version.
>
> cheers
> project-repo
>
> <<A
> `<<A(`
Thanks for the report! This is caused by the recent change to
save/restore here-docment list around command substitutions. In
doing so we must finish existing here-documents prior to restoring
the old here-document list. This is done for new-style command
substitutions but not for old-style.
This patch fixes it by doing it for both.
Reported-by: project-repo <bugs@feusi.co>
Fixes: 51e2d88d6e51 ("parser: Save/restore here-documents in...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
As it is a subshell can execute code that is only meant for the
parent shell when it executes a longjmp that is caught by something
like evalcommand. This patch fixes it by resetting the handler
when entering a subshell.
Reported-by: Martijn Dekker <martijn@inlv.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Due to a double decrement in argstr we may miss field separators
at the end of a word in certain situations.
Reported-by: Martijn Dekker <martijn@inlv.org>
Fixes: 3cd538634f71 ("expand: Do not reprocess data when...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
We set commandname in procargs when we don't have to. This results
in a duplicated output of arg0 when an error occurs.
Reported-by: Olivier Duclos <odc@fastmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The value of REALLY_CLOSED is used to avoid an unnecessary close(2)
call when restoring redirections. However, as it stands it can
remove a close(2) call that's actually needed. This happens when
an enclosed exec(1) command leaves an open file descriptor behind.
This patch fixes this by replacing REALLY_CLOSED with closed_redirs
to track the current status of redirected file descriptors and
leaving redirlist to only handle the previous state of redirected
file descriptors.
Reported-by: Martijn Dekker <martijn@inlv.org>
Fixes: ce0f1900d869 ("[REDIR] Fix redirect restore on saved file...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Building with clang results in some warnings about the use of GNU
old-style field designators:
-----------------------------------------------------------------------
output.c:86:2: warning: use of GNU old-style field designator extension [-Wgnu-designator]
nextc: 0, end: 0, buf: 0, bufsize: OUTBUFSIZ, fd: 1, flags: 0
^~~~~~
.nextc =
...
-----------------------------------------------------------------------
Fix the issue bu using C99 initializers instead.
This should be safe and should not introduce any compatibility problems
as it is done already in other parts of the codebase, like
src/expand.c:ccmatch() and src/parser.c::readtoken1().
Signed-off-by: Antonio Ospite <ao2@ao2.it>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Building with clang results in some warnings about integer values being
added to strings:
-----------------------------------------------------------------------
eval.c:1138:13: warning: adding 'int' to a string does not append to the string [-Wstring-plus-int]
p = " %s" + (1 - sep);
~~~~~~^~~~~~~~~~~
eval.c:1138:13: note: use array indexing to silence this warning
p = " %s" + (1 - sep);
^
& [ ]
1 warning generated.
...
jobs.c:1424:16: warning: adding 'int' to a string does not append to the string [-Wstring-plus-int]
str = "\"}" + !(quoted & 1);
~~~~~~^~~~~~~~~~~~~~~
jobs.c:1424:16: note: use array indexing to silence this warning
str = "\"}" + !(quoted & 1);
^
& [ ]
1 warning generated.
-----------------------------------------------------------------------
While the code itself is fine and the warnings are indeed harmless,
fixing them also makes the semantic more explicit: what it is actually
being increased is the address which points to the start of the string
in order to skip the initial character when some conditions are met.
Signed-off-by: Antonio Ospite <ao2@ao2.it>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
We unconditionally restore the saved status in exitreset, which
is incorrect as we only want to do it for exitcmd and returncmd.
This patch fixes the problem by introducing EXEND.
Reported-by: Martijn Dekker <martijn@inlv.org>
Fixes: da30b4b78769 ("[BUILTIN] Exit without arguments in a trap...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
The following constructs result in ever-increasing memory usage:
while true; do { true; } </dev/null; done
while true; do ( true; ) </dev/null; done
For comparison, bash displays static memory usage in both cases.
This issue was reported for BusyBox ash which is derived from dash:
https://bugs.busybox.net/show_bug.cgi?id=7748
Signed-off-by: Ron Yorston <rmy@frippery.org>
I have simplified evaltree so that it simply sets the stack mark
unconditionally. This allows us to remove the stack marks in the
functions called by evaltree.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
As sigsetmask is set as deprecated in glibc this patch adds the
pragmas to disable the warning in gcc around our one and only use
of sigsetmask.
It also disables it completely for non-gcc compilers and older
gcc compilers as they may generate a warning too.
Reported-by: Antonio Ospite <ao2@ao2.it>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
This patch fixes a typo in evalbltin where warnx was used instead
of sh_warnx.
Reported-by: Antonio Ospite <ao2@ao2.it>
Fixes: 8e43729547b5 ("eval: Report I/O error on stdout")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
When we read the first token in list() we use peektoken instead
of readtoken as the following code needs to use the same token
again. However, this is wrong when we're in a here-document as
it will clobber the saved token without resetting the tokpushback
flag.
This patch fixes it by doing the tokpushback after parseheredoc
and setting lasttoken again if parseheredoc was called.
Reported-by: Ron Yorston <rmy@frippery.org>
Fixes: 7c245aa8ed33 ("[PARSER] Simplify EOF/newline handling in...")
Fixes: ee5cbe9fd6bc ("[SHELL] Optimize dash -c "command" to avoid a fork")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Tested-by: Simon Ser <contact@emersion.fr>
|
|
When we are doing VSLENGTH expansion, the closing brace is currently
not removed in evalvar. This causes the caller argstr to terminate
prematurely as it would interpret the closing brace as one that
belongs to a parameter expansion at the outer level.
This patch fixes it.
Reported-by: Martijn Dekker <martijn@inlv.org>
Fixes: 3cd538634f71 ("expand: Do not reprocess data when...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
It has been reported that
echo test >&$EMPTY_VARIABLE
causes dash to segfault. This is a symptom of the bigger problem
that dash tries to perform pathname expansion as well as field
splitting on the word after >& and <&. This is wrong and this
patch fixes it to use the same expansions as done on a normal
redirection.
Reported-by: Andrej Shadura <andrew.shadura@collabora.co.uk>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Gcc gives a warning about some missing parentheses:
-----------------------------------------------------------------------
eval.c: In function ‘evaltree’:
eval.c:282:15: warning: logical not is only applied to the left hand side of comparison [-Wlogical-not-parentheses]
if (!status == isor || evalskip)
^~
eval.c:282:7: note: add parentheses around left hand side expression to silence this warning
if (!status == isor || evalskip)
^~~~~~~
( )
-----------------------------------------------------------------------
Add the parentheses to silence the warning.
Signed-off-by: Antonio Ospite <ao2@ao2.it>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Enable automake silent rules to make it easier to spot compilation
problems.
Silent rules will be enabled by default, but only if they are available,
in order to keep compatibility with older autotools versions.
Prepend the silent strings also to custom rules.
Signed-off-by: Antonio Ospite <ao2@ao2.it>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|
|
Apply the changes suggested by running autoupdate on the source
repository:
1. Properly quote AC_INIT arguments.
2. Use AC_USE_SYSTEM_EXTENSIONS instead of AC_GNU_SOURCE. The former is
a superset of the latter, and enables more options, see
https://www.gnu.org/software/autoconf/manual/autoconf-2.60/html_node/Posix-Variants.html
Signed-off-by: Antonio Ospite <ao2@ao2.it>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
|