This change fixes the build in setups where there are
no headers defining WIFSTOPPED and WSTOPSIG (where JOBS has to be
set to 0).
This partially reverts 4700fb5be (ash: make dowait() a bit more
readable. Logic is unchanged, 2015-10-09).
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
$ PATH=/extra/path:/usr/sbin:/usr/bin:/sbin:/bin \
> busybox sh -xc 'command -V ls; command -V ls; command -Vp ls; command -vp ls'
+ command -V ls
ls is /bin/ls
+ command -V ls
ls is a tracked alias for /bin/ls
+ command -Vp ls
ls is a tracked alias for (null)
+ command -vp ls
Segmentation fault
describe_command should respect `path' argument. Looking up in the hash table
may gives incorrect index in entry.u.index and finally causes incorrect output
or SIGSEGV.
function old new delta
describe_command 386 313 -73
Signed-off-by: Youfu Zhang <zhangyoufu@gmail.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
When using musl libc glob() a very long string can cause glob() to fail,
which leads to an out of memory error being raised by ash.
This can happen easily if a very long quoted string contains *, even
though no glob expansion should ever be performed on it (since it's
quoted).
Fix this by properly parsing control characters and escaping and only
accept unquoted metacharacters. While we're at it, unify this check for
libc and built-in glob expansion
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
bash has a feature: it restores termios after a successful wait for
a foreground job which had at least one stopped or sigkilled member.
The probable rationale is that SIGSTOP and SIGKILL can preclude task from
properly restoring tty state. Should we do this too?
A reproducer: ^Z an interactive python:
$ python
Python 2.7.12 (...)
>>> ^Z
{ python leaves tty in -icanon -echo state. We do survive that... }
[1]+ Stopped python
{ ...however, next program (python no.2) does not survive it well: }
$ python
Python 2.7.12 (...)
>>> Traceback (most recent call last):
{ above, I typed "qwerty<CR>", but -echo state is still in effect }
File "<stdin>", line 1, in <module>
NameError: name 'qwerty' is not defined
The implementation is modeled on bash code and seems to work.
However, I'm not sure we should do this. For one: what if I'd fg
the stopped python instead? It'll be confused by "restored" tty state.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Mike deleted it:
commit 39456a18a1
Author: Mike Frysinger <vapier@gentoo.org>
Date: Sat Mar 28 12:21:57 2009 +0000
stop lying about [[ test support
probably because it was not properly ifdefed around, and was enabled
even when bash compat is off.
I just tested it - it works:
$ [ *.diff = z.diff ]; echo $?
0
$ [[ *.diff = z.diff ]]; echo $?
1
Of course, not all numerous bash tricks of [[ ]] are implemented...
function old new delta
bltins2 60 72 +12
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Splitting these options makes it self-documenting about what
bash-compatible features we have.
Signed-off-by: Kang-Che Sung <explorer09@gmail.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Splitting these options makes it self-documenting about what
bash-compatible features we have.
Signed-off-by: Kang-Che Sung <explorer09@gmail.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Redundant help texts (one which only repeats the description)
are deleted.
Descriptions and help texts are trimmed.
Some config options are moved, even across menus.
No config option _names_ are changed.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
It's a bit overkill (who would want it off?) but ash already has it
configurable. Let's be symmetric.
Also tweak kbuild logic to use ASH_BUILTIN_ECHO to select echo.o,
not ASH.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Also made it and printf, type and wait builtins optional.
function old new delta
builtin_kill - 323 +323
bltins1 336 348 +12
builtin_type 114 116 +2
------------------------------------------------------------------------------
(add/remove: 1/0 grow/shrink: 2/0 up/down: 337/0) Total: 337 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
commit e19923f665 deleted clearredir()
call in shellexec():
ash: [REDIR] Remove redundant CLOEXEC calls
Upstream commit:
Now that we're marking file descriptors as CLOEXEC in savefd, we no longer
need to close them on exec or in setinputfd.
but it missed one place where we don't set CLOEXEC. Fixing this.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
This would makes all ash options indented inside "ash" in menuconfig.
It appears that menuconfig has a limit at tracking multiple dependency
lines like this (it looks like a "diamond problem" but I'm not sure if
it is):
---ASH <----------
/ \ ASH_OPTIMIZE_FOR_SIZE
!NOMMU <-*----SH_IS_ASH <----[OR] <--ASH_INTERNAL_GLOB
\ / ASH_RANDOM_SUPPORT
---BASH_IS_ASH <-- [...]
The kconfig-language document [1] states that:
> If a menu entry somehow depends on the previous entry, it can be
> made a submenu of it. First, the previous (parent) symbol must be
> part of the dependency list and then one of these two conditions
> must be true:
> - the child entry must become invisible, if the parent is set to 'n'
[BusyBox ash used to satisfy this, but no longer does]
> - the child entry must only be visible, if the parent is visible
[BusyBox ash configs actually satisfy this, but because of
"diamond" above this might not be easily detected]
So I found out a direct workaround: by making ash options explicitly
depend on !NOMMU, we can tell menuconfig that rule 2 above is satisfied
without any more tracking.
---------------------
/ \
!NOMMU <-*-----ASH <-------- \
\ \ \ ASH_OPTIMIZE_FOR_SIZE
*---SH_IS_ASH <---[OR]-[AND] <--ASH_INTERNAL_GLOB
\ / ASH_RANDOM_SUPPORT
--BASH_IS_ASH <- [...]
So all ash options would now be indented under "ash".
[1] "Documentation/kbuild/kconfig-language.txt" in Linux kernel source
Signed-off-by: Kang-Che Sung <explorer09@gmail.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Mention the behavior if user selects CONFIG_SH_IS_ASH but not
CONFIG_ASH. We will be explicit that invocations like "busybox ash"
will not work for such configuration.
Also clarify help text of CONFIG_BASH_IS_* that bash compatibility in
ash is not complete. (It shouldn't be anyway - ash can't support every
bash quirk out there.)
Signed-off-by: Kang-Che Sung <explorer09@gmail.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
The commit 'ash,hush: set exit code 127 in "sh /does/not/exist" case'
only partly implemented the dash commit '[ERROR] Allow the originator
of EXERROR to set the exit status'. This resulted in incorrect error
codes for a syntax error:
$ )
$ echo $?
0
or a redirection error for a special builtin:
$ rm -f xxx
$ eval cat <xxx
$ echo $?
0
Signed-off-by: Ron Yorston <rmy@pobox.com>
Reported-by: Martijn Dekker <martijn@inlv.org>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
The same can be done for msh, but we are probably better off just deleting it
in a next versio or two.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
PIPE_SEQ is used most often, having it zero makes code smaller:
function old new delta
done_word 719 707 -12
parse_stream 2546 2531 -15
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
I thought gcc can detect this itself. It doesn't.
function old new delta
run_list 1030 1021 -9
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Was playing with "sleep 3 | exit 3 & wait %1" and noticed that often
SIGCHLD arrives even before I get to signal masking. Can avoid it in this
case.
function old new delta
wait_for_child_or_signal 228 265 +37
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Also add tests. wait5.tests so far fails (but works for ash and dash).
function old new delta
builtin_wait 305 283 -22
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
It would be nice to provide bash-like "remember las exitcode"
thingy, but it's a bit complex. For now, match ash and dash.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
In this example:
ash -c 'readonly x; echo $(command eval x=2)'
evalstring() is called after forkchild(), which calls popallfiles().
On exception, evalstring() will popfile().
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
It was not properly interruptible, and did not update job status
(the exited processes were still thought of as running).
function old new delta
process_wait_result - 453 +453
wait_for_child_or_signal - 199 +199
run_list 996 1002 +6
checkjobs_and_fg_shell 41 43 +2
builtin_wait 328 215 -113
checkjobs 516 142 -374
------------------------------------------------------------------------------
(add/remove: 2/0 grow/shrink: 2/2 up/down: 660/-487) Total: 173 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit:
Date: Sun, 22 Feb 2009 18:10:01 +0800
[JOBS] Fix dowait signal race
This test program by Alexey Gladkov can cause dash to enter an
infinite loop in waitcmd.
#!/bin/dash
trap "echo TRAP" USR1
stub() {
echo ">>> STUB $1" >&2
sleep $1
echo "<<< STUB $1" >&2
kill -USR1 $$
}
stub 3 &
stub 2 &
until { echo "###"; wait; } do
echo "*** $?"
done
The problem is that if we get a signal after the wait3 system
call has returned but before we get to INTON in dowait, then
we can jump back up to the top and lose the exit status. So
if we then wait for the job that has just exited, then it'll
stay there forever.
I made the original change that caused this bug to fix pretty
much the same bug but in the opposite direction. That is, if
we get a signal after we enter wait3 but before we hit the kernel
then it too can cause the wait to go on forever (assuming the
child doesn't exit).
In fact this is pretty much exactly the scenario that you'll
find in glibc's documentation on pause(). The solution is given
there too, in the form of sigsuspend, which is the only way to
do the check and wait atomically.
So this patch fixes Alexey's race without reintroducing the old
bug by converting the blocking wait3 to a sigsuspend.
In order to do this we need to set a signal handler for SIGCHLD,
so the code has been modified to always do that.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
I failed to reproduce the bug (it requires precise timing), but it seems real.
function old new delta
dowait 284 463 +179
setsignal 301 326 +25
signal_handler 59 76 +17
ash_main 1481 1487 +6
localcmd 350 348 -2
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 4/1 up/down: 227/-2) Total: 225 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit 1:
Date: Sun, 22 Feb 2009 18:16:13 +0800
[SIGNAL] Remove EXSIG
Now that waitcmd no longer uses EXSIG we can remove it.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Upstream commit 2:
Date: Thu, 2 Oct 2014 21:07:55 +0800
[ERROR] Set exitstatus in onint
Currently the exit status when we receive SIGINT is set in evalcommand
which means that it doesn't always get set. For example, if you press
CTRL-C at the prompt of an interactive dash, the exit status is not
set to 130 as it is in many other Bourne shells.
This patch fixes this by moving the setting of the exit status into
onint which also simplifies evalcommand.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Upstream commit 3:
Date: Fri, 3 Oct 2014 14:07:07 +0800
[EVAL] Do not clobber exitstatus in evalcommand
All originators of EXERROR have been setting the exitstatus for
a while now. So it is no longer appropriate to set it explicitly
in evalcommand.
In fact doing so may cause the original exitstatus to be lost.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Last three coomits:
function old new delta
waitcmd 186 224 +38
dowait 276 284 +8
waitforjob 104 107 +3
localcmd 348 350 +2
showjobs 64 61 -3
forkshell 263 260 -3
raise_interrupt 93 67 -26
blocking_wait_with_raise_on_sig 40 - -40
evalcommand 1264 1208 -56
evaltree 809 498 -311
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
This loses an insignificant optimization, but may allow backporting
of some recent-ish dash fixes.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
This commit should have deleted these two statements:
commit c0e007663d
Author: Ron Yorston <rmy@pobox.com>
Date: Thu Oct 29 11:30:55 2015 +0000
ash: simplify EOF/newline handling in list parser
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit:
Date: Wed, 8 Sep 2010 20:07:26 +0800
[EXPAND] Fix ifsfirst/ifslastp leak
As it stands expandarg may return with a non-NULL ifslastp which
then confuses any subsequent ifsbreakup user that doesn't clear
it directly.
What's worse, if we get interrupted before we hit ifsfree in
expandarg we will leak memory.
This patch fixes this by always calling ifsfree in expandarg
thus ensuring that ifslastp is always NULL on the normal path.
It also adds an ifsfree call to the RESET path to ensure that
memory isn't leaked.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Fallout 1:
Date: Mon, 18 Oct 2010 10:55:42 +0800
[EXPAND] Fix ifsfirst/ifslastp leak in casematch
The commit f42e443bb511ed3224f09b4fcf0772438ebdbbfa
[EXPAND] Fix ifsfirst/ifslastp leak
revealed yet another ifsfirst/ifslastp leak in casematch.
Previously it was hidden because ifsfirst/ifslastp was cleared
unconditionally on entry (which caused the leakage of those
entries).
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Fallout 2:
Date: Sun, 28 Nov 2010 21:09:51 +0800
[EXPAND] Free IFS state in evalbackcmd
On Sun, Nov 07, 2010 at 04:04:20PM -0600, Jonathan Nieder wrote:
> Herbert Xu wrote:
> > commit f42e443bb511ed3224f09b4fcf0772438ebdbbfa
> > Author: Herbert Xu <herbert@gondor.apana.org.au>
> > Date: Wed Sep 8 20:07:26 2010 +0800
> >
> > [EXPAND] Fix ifsfirst/ifslastp leak
>
> Another puzzle bisecting to f42e443bb. This one comes from the
> grub-mkconfig script:
>
> $ sh -c 'datadir=/usr/share; pkgdatadir=${datadir}/`cat`' 2>&1 | cat -A
> cat: M-^\^M^F^HM-4^M^F^HM-(^M^F^H: No such file or directory$
> cat: M-(^M^F^H: No such file or directory$
>
> Still reproducible with 016b529. I'll try to find time to look into
> it, but thought you might like to know nevertheless.
This is the symptom of another leak. In this case evalbackcmd
occurs in the middle of an expansion (as it should) but the forked
child never clears the previous IFS state.
This patch adds the missing ifsfree call.
This wasn't as much of a problem as the previously discovered leaks
since all it means is that the child gets to carry around the parent's
expansion state and the child is usually short-lived.
Reported-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Fallout 3:
Date: Tue, 15 Mar 2011 16:01:34 +0800
[EXPAND] Free IFS state after here document expansion
Here's another bug bisecting to f42e443bb ([EXPAND] Fix
ifsfirst/ifslastp leak, 2010-09-08). It was found with the following
test case, based on the configure script for Tracker:
dash -x -c '
<<-_ACEOF
$@
_ACEOF
exec
' - abcdefgh
+
+ exec ?a
exec: 1: : Permission denied
The missing ifsfree call is in expandarg when it returns to openhere
during here document expansion.
Reported-by: Aurelien Jarno <aurel32@debian.org>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
function old new delta
ifsfree - 66 +66
ash_main 1490 1495 +5
argstr 1154 1159 +5
evalcase 275 270 -5
expandarg 972 888 -84
------------------------------------------------------------------------------
(add/remove: 1/0 grow/shrink: 2/2 up/down: 76/-89) Total: -13 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit "[EVAL] Force fork if any trap is set, not just on EXIT"
had a similar code as our fix to that bug.
Eliminate some superficial differences.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
There was a bug in tryexec which bbox had fixed in 2003.
dash had a smaller fix in 2007. Copy it. It is smaller,
although it is also more quirky (requires argv[-1] to exist).
Upstream commit 1:
Date: Mon, 15 Oct 2007 20:24:28 +0800
[EXEC] Fixed execing of scripts with no hash-bang
The function tryexec used the original name instead of the path found through
PATH search. This patch fixes that.
Test case:
trap 'rm -f $TMP' EXIT
TMP=$(tempfile -s nosuchthing)
cat <<- EOF > $TMP
echo OK
EOF
chmod u+x $TMP
cd /
PATH=${TMP%/*} ${TMP##*/}
Old result:
/bin/sh: Can't open filelgY4Fanosuchthing
New result:
OK
Upstream commit 2:
Date: Sun, 23 Dec 2007 11:02:26 +0800
[EVAL] Fix bad pointer arithmetic in evalcommand
dash dies on sparc with a SIGBUS due to an arithmetic error introduced
with commit 03b4958, this patch fixes it.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
function old new delta
evalcommand 1261 1264 +3
dotcmd 321 319 -2
tryexec 115 64 -51
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 1/2 up/down: 3/-53) Total: -50 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit:
Date: Mon, 31 Aug 2009 22:06:41 +1000
[CD] Lookup PWD after going through CDPATH
On Tue, Jul 14, 2009 at 09:39:03PM +0000, Eric Blake wrote:
> For the cd command, POSIX 2008 requires that after all pathnames in CDPATH
> have been tested and failed in step 5, then step 6 interprets the directory
> argument relative to PWD. In other words, this demonstrates a bug:
>
> $ dash -c 'cd /tmp; mkdir -p foo; CDPATH=oops; cd foo; echo $?; pwd'
> cd: 1: can't cd to foo
> 2
> /tmp
>
> while bash gets it correct:
>
> $ bash -c 'cd /tmp; mkdir -p foo; CDPATH=oops; cd foo; echo $?; pwd'
> 0
> /tmp/foo
This patch fixes the problem.
Reported-by: Eric Blake <ebb9@byu.net>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
function old new delta
cdcmd 667 680 +13
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream patch:
Date: Fri, 5 Oct 2007 23:26:45 +0800
[MEMALLOC] Made grabstackblock an inline wrapper for stalloc
The function grabstackblock is identical in semantics to stalloc within its
input constraints.
function old new delta
dotcmd 319 321 +2
grabstackblock 19 5 -14
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit:
Date: Sat, 6 Oct 2007 21:18:58 +0800
[VAR] Remove setvarsafe
The only user of setvarsafe is getopts. However, we can achieve the same
result by pre-setting the value of shellparam.optind.
function old new delta
getoptscmd 614 515 -99
setvarsafe 147 - -147
------------------------------------------------------------------------------
(add/remove: 0/1 grow/shrink: 0/1 up/down: 0/-246) Total: -246 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit:
Date: Thu, 4 Oct 2007 22:20:38 +0800
[PARSER] Size optimisations in parameter expansion parser
Merge flags into subtype.
Do not write subtype out twice.
Add likely flag on ${ vs. $NAME.
Kill unnecessary (and bogus) PEOA check.
function old new delta
readtoken1 2891 2860 -31
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit 1:
Date: Wed, 26 Sep 2007 17:14:16 +0800
[PARSER] Recognise here-doc delimiters terminated by EOF
Previously dash required a <newline> character to be present in order for
a here-document delimiter to be detected. Allowing EOF in the absence of
a <newline> to play the same purpose allows some intuitive scripts to
succeed. POSIX seems to be silence on this so this should be OK.
Test case:
eval 'cat <<- NOT
test
NOT'
echo OK
Old result:
test
NOTOK
New result:
test
OK
Upstream commit 2:
Date: Sat, 20 Oct 2007 18:49:31 +0800
[PARSER] Fix here-doc corruption
The change
[PARSER] Recognise here-doc delimiters terminated by EOF
introduced a regerssion whereby lines starting with eofmark but are not equal
to eofmark would be corrupted. This patch fixes it.
Test case:
cat << _ACEOF
_ASBOX
_ACEOF
Old result:
SASBOX
New result:
_ASBOX
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit:
Date: Thu, 4 Oct 2007 22:15:10 +0800
[PARSER] Fix parsing of ${##1}
Previously dash treated ${##1} as a length operation. This patch fixes that.
Test case:
set -- a
echo ${##1}OK
Old result:
1OK
New result:
OK
This was a real bug in ash (but not in hush).
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit:
Date: Sun, 6 May 2007 19:28:56 +1000
[REDIR] Remove redundant CLOEXEC calls
Now that we're marking file descriptors as CLOEXEC in savefd, we no longer
need to close them on exec or in setinputfd.
function old new delta
ash_main 1478 1492 +14
setinputfile 224 226 +2
readtoken1 2752 2750 -2
shellexec 208 198 -10
clearredir 30 - -30
------------------------------------------------------------------------------
(add/remove: 0/1 grow/shrink: 2/2 up/down: 16/-42) Total: -26 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit:
Date: Sat, 12 May 2007 18:00:57 +1000
[REDIR] Replace copyfd by savefd and use dup2 elsewhere
There are two kinds of users to copyfd, those that want to copy an fd to
an exact value and those that want to move an fd to a value >= 10. The
former can simply use dup2 directly while the latter share a lot of common
code that now constitutes savefd.
This does not change much, just reducing our divergence from dash code.
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstream commit:
Date: Sat, 6 Oct 2007 18:59:31 +0800
[BUILTIN] Treat OPTIND=0 in the same way as OPTIND=1
Previously setting OPTIND to 0 would cause subsequent getopts calls to fail.
This patch makes dash reset the getopts parameters the same way as OPTIND=1.
Both behaviours are allowed by POSIX but other common shells do tolerate this
case.
function old new delta
getoptsreset 24 30 +6
getoptscmd 632 614 -18
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Upstreams commit:
Date: Mon, 8 Oct 2007 21:32:25 +0800
[PARSER] Report substition errors at expansion time
On Wed, Apr 11, 2007 at 01:24:21PM -0700, Micah Cowan wrote:
> This operation fails on Ubuntu:
>
> $ /bin/sh -c 'if false; then d="${foo/bar}"; fi'
> /bin/sh: Syntax error: Bad substitution
>
> When used with other POSIX shells it succeeds. While semantically the
> variable reference ${foo/bar} is not valid, this is not a syntax error
> according to POSIX, and since the variable assignment expression is
> never invoked (because it's within an "if false") it should not be seen
> as an error.
>
> I ran into this because after restarting my system I could no longer log
> in. It turns out that the problem was (a) I had edited .gnomerc to
> source my .bashrc file so that my environment would be set properly, and
> (b) I had added some new code to my .bashrc WITHIN A CHECK FOR BASH!
> that used bash's ${var/match/sub} feature. Even though this code was
> within a "case $BASH_VERSION; in *[0-9]*) ... esac (so dash would never
> execute it since that variable is not set), it still caused dash to
> throw up.
>
> FYI, some relevant details from POSIX:
>
> Section 2.3, Token Recognition:
>
> 5. If the current character is an unquoted '$' or '`', the shell shall
> identify the start of any candidates for parameter expansion ( Parameter
> Expansion), command substitution ( Command Substitution), or arithmetic
> expansion ( Arithmetic Expansion) from their introductory unquoted
> character sequences: '$' or "${", "$(" or '`', and "$((", respectively.
> The shell shall read sufficient input to determine the end of the unit
> to be expanded (as explained in the cited sections).
>
> Section 2.6.2, Parameter Expansion:
>
> The format for parameter expansion is as follows:
>
> ${expression}
>
> where expression consists of all characters until the matching '}'. Any
> '}' escaped by a backslash or within a quoted string, and characters in
> embedded arithmetic expansions, command substitutions, and variable
> expansions, shall not be examined in determining the matching '}'.
> [...]
>
> 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 followed by a character that could be interpreted as
> part of the name. The matching closing brace shall be determined by
> counting brace levels, skipping over enclosed quoted strings, and
> command substitutions.
> ---
> In addition to bash I've checked Solaris /bin/sh and ksh and they don't
> report an error.
>
> -----
> Micah Cowan:
>
> The applicable portion of POSIX is in XCU 2.10.1:
>
> "The WORD tokens shall have the word expansion rules applied to them
> immediately before the associated command is executed, not at the time
> the command is parsed."
>
> This seems fairly clear to me.
This patch moves the error detection to expansion time.
Test case:
if false; then
echo ${a!7}
fi
echo OK
Old result:
dash: Syntax error: Bad substitution
New result:
OK
function old new delta
evalvar 574 585 +11
readtoken1 2763 2750 -13
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
