The bugfix for BUG_CMDSPASGN backported in commit fae8862c caused
two regressions with the += operator:
1. The += operator did not append to variables. Reproducer:
$ integer foo=3
$ foo+=2 command eval 'echo $foo'
2
2. The += operator ignored the readonly attribute, modifying readonly
variables in the same manner as above. Reproducer
$ readonly bar=str
$ bar+=ing command eval 'echo $bar'
ing
Both of the regressions above were caused by nv_putval() failing to
clone the variable from the previous scope into the invocation-local
scope. As a result, 'foo+=2' was effectively 0 + 2 (since ksh didn't
clone 3). The first regression was noticed during the development of
ksh93v-, so to fix both bugs I've backported the bugfix for the
regression from the ksh93v- 2013-10-10 alpha version:
https://www.mail-archive.com/ast-users@lists.research.att.com/msg00369.html
src/cmd/ksh93/sh/name.c:
- To fix both of the bugs above, find the variable to modify with
nv_search(), then clone it into the invocation local scope. To
fix the readonly bug as well, this is done before the NV_RDONLY
check (otherwise np will be missing that attribute and be
incorrectly modified in the invocation-local scope).
- Update a nearby comment describing what sh_assignok() does (per this
comment: https://github.com/ksh93/ksh/pull/249#issuecomment-811381759)
src/cmd/ksh93/tests/builtins.sh:
- Add regression tests for both of the now fixed regressions,
loosely based on the regression tests in ksh93v-.
src/cmd/ksh93/tests/readonly.sh:
- Use a 'ulimit --cpu' as a workaround to close down hung processes
that might be caused due to a couple of known bugs (recursion and
type variable function)
Discussion: https://github.com/ksh93/ksh/issues/264
- Adjust tests so xtrace can be used
- Use integer n within for loop
The recursion level for arithmetic expressions is kept track of in
a static 'level' variable in streval.c. It is reset when arithmetic
expressions throw an error.
But an error for an arithmetic expression may also occur elsewhere
-- at least in one case: when an arithmetic expression attempts to
change a read-only variable. In that case, the recursion level is
never reset because that code does not have access to the static
'level' variable.
If many such conditions occur (as in the new readonly.sh regression
tests), an arithmetic command like 'i++' may eventually fail with a
'recursion too deep' error.
To mitigate the problem, MAXLEVEL in streval.c was changed from 9
to 1024 in 264ba48b (as in the ksh 93v- beta). This commit leaves
that increase, but adds a proper fix.
src/cmd/ksh93/include/defs.h:
- Add global sh.arithrecursion (a.k.a. shp->arithrecursion)
variable to keep track of the arithmetic recursion level,
replacing the static 'level' variable in streval.c.
src/cmd/ksh93/sh/xec.c: sh_exec():
- Reset sh.arithrecursion before starting a new simple command
(TCOM), a new subshell with parentheses (TPAR), a new pipe
(TFIL), or a new [[ ... ]] command (TTST). These are the same
places where 'echeck' is set to 1 for --errexit and ERR trap
checks, so it should cover everything.
src/cmd/ksh93/sh/streval.c:
- Change all uses of 'level' to sh.arithrecursion.
- _seterror, aritherror(): No longer bother to reset the level
to zero here; xec.c should have this covered for all cases now.
src/cmd/ksh93/tests/arith.sh:
- Add tests for main shell and subshell.
One area where readonly is still ineffective is the local
environment list for a command (preceding assignments) if that
command is not executed using exec(3) after fork(2). Builtin
commands are one example. The following succeeds but should fail:
(readonly v=1; v=2 true) # succeeds, but should fail
If the shell is compiled with SHOPT_SPAWN (the default) then this
also applies to external commands invoked with sh_ntfork():
(readonly v=1; v=2 env) # succeeds if SHOPT_SPAWN
This presents to the user as inconsitent behaviour because external
commands may be fork()ed under certain circumstances but not
others, depending on complex optimisations. One example is:
$ ksh -c 'readonly v=1; v=2 env'
ksh: v: is read only
$ ksh -c 'readonly v=1; v=2 env; :'
(bad: environment list is output, including 'v=2')
In the first command above, where 'v2=env' is the last command in
the -c script, the optimisation skips creating a scope and assigns
the environment list in the current scope.
src/cmd/ksh93/sh/name.c: nv_setlist():
- Add check for readonly. This requires searching for the variable
in the main tree using nv_search() before a locally scoped one is
added using nv_open(). Since nv_search() only works with plain
variable names, temporarily end the string at '='.
src/cmd/ksh93/tests/readonly.sh:
- Add version check and fork the test command substitution subshell
on older versions that would otherwise abort the tests due to the
combination of an excessively low arithmetic recursion tolerance
and a bug that sometimes fails to restore the shell's arithmetic
recursion level.
Since f207cd57, sh_ntfork() is never called if job.jobcontrol is
set (i.e. if job control is active on an interactive shell), so the
code that is only run if job.jobcontrol is set should be removed.
src/cmd/ksh93/sh/xec.c:
- Remove spawnveg() define that is unused as of 7b0e0776.
- sh_exec(): Simplify SHOPT_SPAWN preprocessor logic. As sh_fork()
never returns a negative value, only run the parent<0 check after
running sh_ntfork() -- that check already didn't happen when
compiling ksh with SHOPT_SPAWN disabled.
- sh_ntfork(): Remove signal and terminal handling (with race
condition) that was only run with job.jobcontrol set.
src/cmd/ksh93/sh/args.c: sh_argopts():
- Remove special-casing for --posix (see also data/builtins.c) and
move the case -5: to the case ':' instead, so this option is
handled like all other long options. This change fixes two bugs:
1. 'set --posix' had no effect on the letoctal or braceexpand
options. Reproducer:
$ set --posix
$ [[ -o braceexpand ]]; echo $?
0
$ [[ -o letoctal ]]; echo $?
1
2. 'ksh --posix' could not run scripts correctly because it
wrongly enabled '-c'. Reproducer:
$ ksh --posix < <(echo 'exit 0')
ksh: -c requires argument
Usage: ksh [--posix] [arg ...]
Help: ksh [ --help | --man ] 2>&1
- Don't allow 'set --default' to unset the restricted option.
src/cmd/ksh93/tests/options.sh:
- Add regression tests for the bugs described above, using -o posix
and --posix.
src/cmd/ksh93/tests/restricted.sh:
- Add a regression test for 'set --default' in rksh.
Co-authored-by: Martijn Dekker <martijn@inlv.org>
Using the stack makes it impossible for future buffer overflows to
occur. It also simplifies fmttoken() by eliminating the need to
declare a local buffer and pass a pointer to that as an argument.
For info: man src/lib/libast/man/stak.3
src/lib/libast/tm/tmlocale.c:
- Load the locale set by LC_TIME or LC_ALL if it hasn't been loaded
before or if it was loaded previously but isn't the current locale.
src/cmd/ksh93/tests/locale.sh:
- Add a regression test using the nl_NL.UTF-8 and ja_JP.UTF-8 locales.
Fixes: https://github.com/ksh93/ksh/issues/261
fmttoken() needs a minimal char[4] token buffer passed to it.
Originally reported by: Jakub Wilk <jwilk@jwilk.net>
Original bug report: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=879464
The following code lines from fmttoken() yield a n=3 for SYMSEMI as
n=1 from the start, e.g. 'for <>;'.
case SYMSEMI:
if(tok[0]=='<')
tok[n++] = '>';
sym = ';';
break;
default:
sym = 0;
}
tok[n++] = sym;
}
tok[n] = 0;
n[0]='<'
n[1]='>'
n[2]=';'
n[3]=0 # <-- BUFFER overflow as the passed character buffers have a size of 3
src/cmd/ksh93/sh/lex.c:
- DBUG: sh_lex(): Adjust char tokstr[3] to char tokstr[4]
- sh_syntax(): Adjust char tokbuf[3] to char tokbuf[4]
Many of these changes are minor typo fixes. The other changes
(which are mostly compiler warning fixes) are:
NEWS:
- The --globcasedetect shell option works on older Linux kernels
when used with FAT32/VFAT file systems, so remove the note about
it only working with 5.2+ kernels.
src/cmd/ksh93/COMPATIBILITY:
- Update the documentation on function scoping with an addition
from ksh93v- (this does apply to ksh93u+).
src/cmd/ksh93/edit/emacs.c:
- Check for '_AST_ksh_release', not 'AST_ksh_release'.
src/cmd/INIT/mamake.c,
src/cmd/INIT/ratz.c,
src/cmd/INIT/release.c,
src/cmd/builtin/pty.c:
- Add more uses of UNREACHABLE() and noreturn, this time for the
build system and pty.
src/cmd/builtin/pty.c,
src/cmd/builtin/array.c,
src/cmd/ksh93/sh/name.c,
src/cmd/ksh93/sh/nvtype.c,
src/cmd/ksh93/sh/suid_exec.c:
- Fix six -Wunused-variable warnings (the name.c nv_arrayptr()
fixes are also in ksh93v-).
- Remove the unused 'tableval' function to fix a -Wunused-function
warning.
src/cmd/ksh93/sh/lex.c:
- Remove unused 'SHOPT_DOS' code, which isn't enabled anywhere.
https://github.com/att/ast/issues/272#issuecomment-354363112
src/cmd/ksh93/bltins/misc.c,
src/cmd/ksh93/bltins/trap.c,
src/cmd/ksh93/bltins/typeset.c:
- Add dictionary generator function declarations for former
aliases that are now builtins (re: 1fbbeaa1, ef1621c1, 3ba4900e).
- For consistency with the rest of the codebase, use '(void)'
instead of '()' for print_cpu_times.
src/cmd/ksh93/sh/init.c,
src/lib/libast/path/pathshell.c:
- Move the otherwise unused EXE macro to pathshell() and only
search for 'sh.exe' on Windows.
src/cmd/ksh93/sh/xec.c,
src/lib/libast/include/ast.h:
- Add an empty definition for inline when compiling with C89.
This allows the timeval_to_double() function to be inlined.
src/cmd/ksh93/include/shlex.h:
- Remove the unused 'PIPESYM2' macro.
src/cmd/ksh93/tests/pty.sh:
- Add '# err_exit #' to count the regression test added in
commit 113a9392.
src/lib/libast/disc/sfdcdio.c:
- Move diordwr, dioread, diowrite and dioexcept behind
'#ifdef F_DIOINFO' to fix one -Wunused-variable warning and
multiple -Wunused-function warnings (sfdcdio() only uses these
functions when F_DIOINFO is defined).
src/lib/libast/string/fmtdev.c:
- Fix two -Wimplicit-function-declaration warnings on Linux by
including sys/sysmacros.h in fmtdev().
There's an annoying inconsistency in error messages if ksh is
compiled with SHOPT_SPAWN. One way to trigger it:
$ /usr/local/bin/ksh -c '/tmp/nonexistent'
/usr/local/bin/ksh: /tmp/nonexistent: not found
$ /usr/local/bin/ksh -c '/tmp/nonexistent; :'
/usr/local/bin/ksh: /tmp/nonexistent: not found [No such file or directory]
In the first variant, as an optimisation, ksh went straight to
exec'ing the command without forking first. In the second variant,
sh_ntfork() was used.
The first variant is done in path_exec(), path.c, line 1049:
errormsg(SH_DICT,ERROR_exit(ERROR_NOENT),e_found,arg0);
The second one is in sh_ntfork(), xec.c, line 3654:
errormsg(SH_DICT,ERROR_system(ERROR_NOENT),e_found+4);
In both cases, the e_found message is only used if errno==ENOENT,
so the extra '[No such file or directory]' message generated by
ERROR_system() is pointless as that will never change for that
message.
src/cmd/ksh93/sh/xec.c: sh_ntfork():
- Use ERROR_exit() instead of ERROR_system() for the e_found
message to avoid the superfluous addition.
If a system administrator prefixes /opt/ast/bin to the path and
then invokes the shell in restricted mode, they clearly intend for
the user to run those AST utilities.
Similarly, if a system administrator sets a PATH for a restricted
shell that includes libraries listed in the .paths file, they must
have intended for the user to use those loadable built-ins, as they
will be associated with the pathnames of their respective
libraries. Since the user cannot change PATH or use the builtin
command, they still cannot load just any built-in they choose.
src/cmd/ksh93/sh/path.c:
- Remove SH_RESTRICTED check when handling path-bound builtins
or dynamic libaries containining builtins in $PATH.
src/cmd/ksh93/tests/builtins.sh:
- Add test verifying a restricted user can use /opt/ast/bin/cat
via a PATH search.
Progresses: https://github.com/ksh93/ksh/issues/138
This commit fixes BUG_CSUBSTDO, which could break stdout inside of
non-forking command substitutions. The breakage only occurred when
stdout was closed outside of the command substitution and a file
descriptor other than stdout was redirected in the command substitution
(such as stderr). Thanks to the ast-open-history repo, I was able to
identify and backport the bugfix from ksh93v- 2012-08-24.
This backport may fix other bugs as well. On 93v- 2012-08-24 it
fixed the regression below, though it was not triggered on 93u+(m).
src/cmd/ksh93/tests/heredoc.sh
487 print foo > $tmp/foofile
488 x=$( $SHELL 2> /dev/null 'read <<< $(<'"$tmp"'/foofile) 2> /dev/null;print -r "$REPLY"')
489 [[ $x == foo ]] || err_exit '<<< $(<file) not working'
src/cmd/ksh93/sh/io.c: sh_open():
- If the just-opened file descriptor exists in sftable and is
flagged with SF_STRING (as in non-forking command substitutions,
among other situations), then move the file descriptor to a
number >= 10.
src/cmd/ksh93/tests/io.sh:
- Add a regression test for BUG_CSUBSTDO, adapted from the one in
modernish.
The current version of 93u+m does not have proper support for the
LC_TIME variable. Setting LC_TIME has no effect on printf %T, and
if the locale is invalid no error message is shown:
$ LC_TIME=ja_JP.UTF-8
$ printf '%T\n' now
Wed Apr 7 15:18:13 PDT 2021
$ LC_TIME=invalid.locale
$ # No error message
src/cmd/ksh93/data/variables.c,
src/cmd/ksh93/include/variables.h,
src/cmd/ksh93/sh/init.c:
- Add support for the $LC_TIME variable. ksh93v- attempted to add
support for LC_TIME, but the patch from that version was extended
because the variable still didn't function correctly.
src/cmd/ksh93/tests/variables.sh:
- Add LC_TIME to the regression tests for LC_* variables.
$ /usr/local/bin/ksh -c 'readonly v=1; export v'
/usr/local/bin/ksh: export: v: is read only
Every POSIX shell (even zsh, as of 5.8) allows this. So did ksh,
until the referenced commit.
src/cmd/ksh93/bltins/typeset.c: setall():
- Allow setting attributes on a readonly variable if any of
NV_ASSIGN (== NV_NOFREE), NV_EXPORT or NV_RDONLY are the only
flag bits that are set. This allows readonly, export, typeset -r,
typeset -x, and typeset -rx on variable arguments without an
assignment. Note that NV_ASSIGN is set for the first variable
argument even though it is not an assignment, so we must allow
it. The logic (or lack thereof) of that is yet to be worked out.
src/cmd/ksh93/tests/readonly.sh:
- Tests.
Resolves: https://github.com/ksh93/ksh/issues/258
This experiment, the initialisation of which was disabled with '#if
0', defines a bunch of integer type commands as special builtins.
Most are boring as they define variables just like normal integers:
pid_t, size_t, etc.
One is interesting: mode_t is a type that automatically converts
from a octal permission bits (e.g. 755) to a mode string like
u+rwx,g+rw,o+rw. That's not a compelling enough use case to
permanently define a special and immutable builtin though.
stat_t is odd: it takes a file name as an argument and fills the
variable with stat information, but it is base64 encoded binary
data and there doesn't seem to be anything that can parse it.
Anyway, none of this is going to be enabled, so we should get rid.
This is an update to one of Red Hat's patches. The strdup change is
from CentOS:
https://git.centos.org/rpms/ksh/blob/c8s/f/SOURCES/ksh-20120801-annocheck.patch
The reason why gcc (and also clang) optimize out the null check is
because the glibc string.h header gives 's' a nonnull attribute (in
other words, this is a glibc compatibility bug, not a compiler bug).
Clang gives the following informative warning when compiling strdup:
> /home/johno/GitRepos/KornShell/ksh/src/lib/libast/string/strdup.c:66:10: warning: nonnull parameter 's' will evaluate to 'true' on
> return (s && (t = oldof(0, char, n = strlen(s) + 1, 0))) ? (char*)memcpy(t, s, n) : (char*)0;
> ^ ~~
> /usr/include/string.h:172:35: note: declared 'nonnull' here
> __THROW __attribute_malloc__ __nonnull ((1));
> ^
> /usr/include/sys/cdefs.h:303:44: note: expanded from macro '__nonnull'
> # define __nonnull(params) __attribute__ ((__nonnull__ params))
The proper fix is to rename the function in strdup.c to
'_ast_strdup'. This avoids the string.h conflict and fixes the Red
Hat bug. I've also made a similar change to getopt.c, since clang
was throwing a nonnull warning there as well.
src/lib/libast/features/map.c (which generates FEATURE/map which is
indirectly included by everything) is updated to always map getopt
to _ast_getopt and strdup to _ast_strdup.
Renamed: src/cmd/INIT/cc.linux.i386 -> src/cmd/INIT/cc.linux
This ensures that architectures like ARM also use the default Linux
wrapper. This is needed because they may need -D_LARGEFILE64_SOURCE
to compile correctly.
On ARM processors, this fixes at least this regression:
io.sh[243]: long seek not working
Resolves: https://github.com/ksh93/ksh/issues/253
The typecast fix was insufficient, avoiding the crash only when
compiling with optimisation disabled. The real problem is that
put_lineno() was passed a misaligned pointer, and that the value
didn't actually contain a double but a string. The bug occurred
when restoring the LINENO value upon exiting a virtual subshell.
Thanks to Harald van Dijk for figuring out the fix.
src/cmd/ksh93/sh/subshell.c: nv_restore():
- When restoring a special variable as defined by nv_cover(),
do not pass either the np->nvflag bits or NV_NOFREE. Why?
* The np->nvflag bits are not needed. They are also harmful
because they may include the NV_INTEGER bit. This is set
when the value is numeric. However, nv_getval() always
returns the value in string form, converting it if it is
numeric. So the NV_INTEGER flag should never be passed
to nv_putval() when it uses the result of nv_getval().
* According to nval.3, the NV_NOFREE flag stops nv_putval() from
creating a copy of the value. But this should be unnecessary
because the earlier _nv_unset(mp,NV_RDONLY|NV_CLONE) should
ensure there is no previous value. In addition, the NV_NOFREE
flag triggered another bug that caused the value of SECONDS to
be corrupted upon restoring it when exiting a virtual subshell.
- When restoring a regular variable, copy the entire nvalue union
and not just the 'cp' member. In practice this worked because
no current member of the nvalue union is larger than a pointer.
However, there is no guarantee it will stay that way.
src/cmd/ksh93/tests/leaks.sh:
- Add disabled test for a memory leak that was discovered in the
course of dealing with this bug. The fix doesn't introduce or
influence it. It will have to be dealt with later.
src/cmd/ksh93/tests/locale.sh:
- Add test for restoring locale on leaving virtual subshell.
https://github.com/ksh93/ksh/issues/253#issuecomment-815290154
src/cmd/ksh93/tests/variables.sh:
- Test against corruption of SECONDS on leaving virtual subshell.
https://github.com/ksh93/ksh/issues/253#issuecomment-815191052
Co-authored-by: Harald van Dijk <harald@gigawatt.nl>
Progresses: https://github.com/ksh93/ksh/issues/253
On Ubuntu arm7, two variables.sh regression tests crashed with a
bus error (SIGBUS) in init.c on line 720 while testing $LINENO:
707 static void put_lineno(Namval_t* np,const char *val,int flags,Namfun_t *fp)
708 {
709 register long n;
710 Shell_t *shp = sh_getinterp();
711 if(!val)
712 {
713 fp = nv_stack(np, NIL(Namfun_t*));
714 if(fp && !fp->nofree)
715 free((void*)fp);
716 _nv_unset(np,NV_RDONLY);
717 return;
718 }
719 if(flags&NV_INTEGER)
720 n = *(double*)val;
721 else
722 n = sh_arith(shp,val);
723 shp->st.firstline += nget_lineno(np,fp)+1-n;
724 }
Apparently, gcc on arm7 doesn't like the implicit typecast from
double to long.
Those three $LINENO discipline functions are generally a mess of
implicit typecasts between Sfdouble_t, double, long and int.
Line numbers are internally stored as int. The discipline functions
need to use Sfdouble_t for API compatibility.
src/cmd/ksh93/sh/init.c: nget_lineno(), put_lineno(), get_lineno():
- Get rid of unnecessary implicit typecasts by adjusting the types
of local variables.
- Make the typecasts that are done explicit.
Progresses: https://github.com/ksh93/ksh/issues/253
On some systems (such as Ubuntu on ARM), the output of `file`
contains a build hash, such as:
SomeExecutable: ELF 32-bit LSB shared object, ARM, EABI5
version 1 (SYSV), dynamically linked, interpreter
/lib/ld-linux-armhf.so.3, for GNU/Linux 3.2.0,
BuildID[sha1]=8934dd61657aac875c190535066466849687a56b,
not stripped
This build hash can contain the string '64', which caused package
to wrongly detect a 64-bit architecture.
bin/package, src/cmd/INIT/package.sh:
- Export LC_ALL=C to ensure 'file' output in English.
- To detect a 64-bit architecture, require the string "64-bit", "64
bit" or "64bit" in 'file' output. The letters 'i' and 't' cannot
occur in a hexadecimal hash, so hopefully that is safe enough. It
is impossible to make this method completely safe, so in the long
term it should be replaced.
Progresses: https://github.com/ksh93/ksh/issues/253
These fixes are applied rather blindly as no one has yet managed to
understand the almost entirely uncommented arrays and variables
handling code (arrays.c, name.c, nvdisc.c, nvtree.c, nvtype.c).
Hopefully we'll figure all that out at some point. In the meantime
these backported fixes appear to work fine, and these bugs impact
the usability of 'enum', so I'm just going to have to violate my
own policy and backport these fixes without understanding them.
Thanks to @JohnoKing for putting in a lot of work tracing these.
Further discussion at: https://github.com/ksh93/ksh/issues/87
src/cmd/ksh93/sh/array.c:
- nv_arraysettype():
* Further simplify the function. After my initial simplification
of it (re: 5491fe97), I don't believe there's actually a need
to save a duplicate copy of the value. Use the pointer returned
by nv_getval() directly to restore the value.
* Cope with a null value (nv_getval() returning a NULL pointer).
This is needed for compatibility with the backported fix in
nvtype.c (below).
- array_putval(): If the array's value pointer (up->cp) is a
pointer to the empty string, it is set to NULL before calling
nv_putv() to prevent an empty string from being deleted. Backport
a fix from 93v- that restores the pointer to the empty string if
the NV_NOFREE attribute is set. Removing it somehow causes these
regressions:
enum.sh[86]: ${array[@]} doesn't yield all values for
associative enum arrays (expected 'green blue blue red
yellow green red orange'; got 'green blue blue yellow
green orange')
enum.sh[94]: unsetting associative enum array does not work
(got 'Color_t -A Colors=([foo]=red [rood]=red)')
enum.sh[116]: assigning first enum element to indexed array
failed (expected 'red red'; got 'BUG BUG')
- nv_associative(): Do not increase the 'nelem' (number of
elements) value of the array's 'header' struct if the array is
associative and of an enum type. The original 93v- fix only
checked for the NV_INTEGER attribute, but backporting that caused
several regressions. Using a debug output command I've determined
that the exact value of 'type' is somehow consistently set to
0x26 if the array is associative and of an enum type, which is
NV_INTEGER | NV_LTOU | NV_RJUST as defined in include/nval.h. I
cannot find where/how that value is determined. In any case this
fix, based on but more specific than the 93v- one, appears to
work fine. Removing it somehow causes this regression:
enum.sh[94]: unsetting associative enum array does not work
(got 'Color_t -A Colors=()')
src/cmd/ksh93/sh/nvtype.c: nv_settype():
- Another fix backported from 93v-. If the variable is an array,
also set the type of element 0 of that array using a call to
nv_arraysettype(). The value may be null. Removing this somehow
causes this regression:
enum.sh[94]: unsetting associative enum array does not work
(got 'Color_t -A Colors=()')
src/cmd/ksh93/tests/enum.sh:
- Add tests for all the bugs fixed here, plus some hypothetical
bugs (e.g., do the same tests for indexed enum type arrays as for
associative enum type arrays, even though indexed enum type
arrays didn't have all the same problems).
Co-authored-by: Johnothan King <johnothanking@protonmail.com>
Resolves: https://github.com/ksh93/ksh/issues/87
Simple reproducer:
set -A arr a b c d; : ${arr[1..2]}; unset arr[1]; echo ${arr[@]}
Output:
a
Expected output:
a c d
The ${arr[1..2]} expansion broke the subsequent 'unset' command
so that it unsets element 1 and on, instead of only 1.
This regression was introduced in nv_endsubscript() on 2009-07-31:
https://github.com/ksh93/ast-open-history/commit/c47896b4/src/cmd/ksh93/sh/array.c
That change checks for the ARRAY_SCAN attribute which enables
processing ranges of array elements instead of single array
elements, and restores it after. That restore is evidently not
correct as it causes the subsequent unset command to malfunction.
If we revert that change, the bug disappears and the regression
tests show no failures. However, I don't know what this was meant
to accomplish and what other bug we might introduce by reverting
this. However, no corresponding regression test was added along
with the 2009-07-31 change, nor is there any corresponding message
in the changelog. So this looks to be one of those mystery changes
that we'll never know the reason for.
Since we currently have proof that this change causes breakage and
no evidence that it fixes anything, I'll go ahead and revert it
(and add a regression test, of course). If that causes another
regression, hopefully someone will find it at some point.
src/cmd/ksh93/sh/array.c: nv_endsubscript():
- Revert the 2009-07-31 change that saves/restores the ARRAY_SCAN
attribute.
- Keep the 'ap' pointer as it is now used by newer code. Move the
declaration up to the beginning of the block, as is customary.
src/cmd/ksh93/sh/init.c:
- Cosmetic change: remove an unused array_scan() macro that I found
when grepping the code for ARRAY_SCAN. The macro was introduced
in version 2001-06-01 but the code that used it was replaced in
version 2001-07-04, without removing the macro itself.
Resolves: https://github.com/ksh93/ksh/issues/254
To set a window title in bash and zsh, the $PS1 prompt can be set
with the title placed between $'\E]0;' and $'\a':
set -o emacs # Or vi mode
typeset -A fmt=(
[start_title]=$'\E]0;'
[end_title]=$'\a'
)
PS1="${fmt[start_title]}$(hostname): $(uname)${fmt[end_title]}\$ "
This also works in ksh unless the shell receives SIGWINCH. With a
$PS1 that sets a window title, the prompt breaks until two
interrupts are received. This is caused by ed_setup() skipping
$'\a' (the bell character) when setting up the e_prompt buffer
which is an edited version of the final line of the PS1 prompt for
use when redrawing the command line.
One fix would be to avoid cutting out the bell character. But if
the prompt contains a bell, we only want the terminal to beep when
a new prompt is printed, and not upon refreshing the command line,
e.g. when receiving SIGWINCH or pressing Ctrl+L.
To avoid the problem, this commit adds code that cuts out sequences
of the form ESC ] <number> ; <text> BELL from the prompt redraw
buffer altogether. They are not needed there because these
sequences will already have taken effect when the full prompt was
printed by io_prompt().
This commit also adds a tweak that should improve the recognition
of other escape sequences to count their length.
src/cmd/ksh93/edit/edit.c: ed_setup():
- When preparing the e_prompt buffer, cut out dtterm/xterm
Operating System Commands that set window/icon title, etc.
See: https://invisible-island.net/xterm/ctlseqs/ctlseqs.html
- When counting the length of escape sequences in that part of PS1,
try to recognize some more types of sequences. These changes are
part of a ksh2020 patch: https://github.com/att/ast/issues/399
src/cmd/ksh93/sh.1:
- Document that any '!' in escape sequences in the PS1 prompt needs
to be changed to '!!'. To avoid breaking compatibility, this
requirement is documented instead of backporting the changes to
io_prompt() from https://github.com/att/ast/issues/399 which try
to remove that requirement for specific escape sequences.
Co-authored-by: Martijn Dekker <martijn@inlv.org>
Ksh currently restricts readonly scalar variables from having their
values directly changed via a value assignment. However, since ksh
allows variable attributes to be altered, the variable's value can
be indirectly altered. For instance, if TMOUT=900 (for a 15 minute
idle timeout) was set to readonly, all that is needed to alter the
value of TMOUT from 900 to 0 is to issue 'typeset -R1 TMOUT',
perhaps followed by a 'typeset -i TMOUT' to turn off the shell's
timeout value.
In addition, there are problems with arrays. The following is
incorrectly allowed:
typeset -a arr=((a b c) 1)
readonly arr
arr[0][1]=d
arr=(alphas=(a b c);name=x)
readonly arr.alphas
arr.alphas[1]=([b]=5)
arr=(alphas=(a b c);name=x)
readonly arr.alphas
arr.alphas[1]=(b)
typeset -C arr=(typeset -r -a alphas=(a b c);name=x)
arr.alphas[1]=()
src/cmd/ksh93/bltins/typeset.c: setall():
- Relocate readonly attribute check higher up the code and widen
its application to issue an error message if the pre-existing
name-pair has the readonly bit flag set.
- To avoid compatibility problems, don't check for readonly if
NV_RDONLY is the only attribute set (ignoring NV_NOFREE). This
allows 'readonly foo; readonly foo' to keep working.
src/cmd/ksh93/sh/array.c: nv_endsubscript():
- Apply a readonly flag check when an array subscript or append
assignment occurs, but allow type variables (typeset -T) as they
utilize '-r' for 'required' sub-variables.
src/cmd/ksh93/tests/readonly.sh:
- New file. Create readonly tests that validate the warning message
and validate that the readonly variable did not change.
src/cmd/ksh93/sh/streval.c:
- Bump MAXLEVEL from 9 to 1024 as a workaround for arithmetic
expansion, avoiding a spurious error about too much recursion
when the readonly.sh tests are run. This change is backported
from ksh 93v-.
TODO: debug a spurious increase in arithmetic recursion level
variable when readonly.sh tests with 'typeset -i' are run.
That is a different bug for a different commit.
Co-authored-by: Martijn Dekker <martijn@inlv.org>
This commit adds an UNREACHABLE() macro that expands to either the
__builtin_unreachable() compiler builtin (for release builds) or
abort(3) (for development builds). This is used to mark code paths
that are never to be reached.
It also adds the 'noreturn' attribute to functions that never
return: path_exec(), sh_done() and sh_syntax(). The UNREACHABLE()
macro is not added after calling these.
The purpose of these is:
* to slightly improve GCC/Clang compiler optimizations;
* to fix a few compiler warnings;
* to add code clarity.
Changes of note:
src/cmd/ksh93/sh/io.c: outexcept():
- Avoid using __builtin_unreachable() here since errormsg can
return despite using ERROR_system(1), as shp->jmplist->mode is
temporarily set to 0. See: https://github.com/att/ast/issues/1336
src/cmd/ksh93/tests/io.sh:
- Add a regression test for the ksh2020 bug referenced above.
src/lib/libast/features/common:
- Detect the existence of either the C11 stdnoreturn.h header or
the GCC noreturn attribute, preferring the former when available.
- Test for the existence of __builtin_unreachable(). Use it for
release builds. On development builds, use abort() instead, which
crahses reliably for debugging when unreachable code is reached.
Co-authored-by: Martijn Dekker <martijn@inlv.org>
This commit fixes a bug in the ksh uname builtin's -d option that could
change the output of -o (I was only able to reproduce this on Linux):
$ builtin uname
$ uname -o
GNU/Linux
$ uname -d
(none)
$ uname -o
(none)
I identified this patch from ksh2020 as a fix for this bug:
<https://github.com/att/ast/pull/1187>
The linked patch was meant to fix a crash in 'uname -d', although I've
had no luck reproducing it: <https://github.com/att/ast/issues/1184>
src/lib/libcmd/uname.c:
- Pass correct buffer to getdomainname() while executing uname -d.
src/cmd/ksh93/tests/builtins.sh:
- Add a regression test for the reported 'uname -d' crash.
- Add a regression test for the output of 'uname -o' after 'uname -d'.
- To handle potential crashes when running the regression tests in older
versions of ksh, fork the command substitutions that run 'uname -d'.
This bug was first reported at <https://github.com/att/ast/issues/8>.
The 'cd' command currently takes the value of $OLDPWD from the
wrong scope. In the following example 'cd -' will change the
directory to /bin instead of /tmp:
$ OLDPWD=/bin ksh93 -c 'OLDPWD=/tmp cd -'
/bin
src/cmd/ksh93/bltins/cd_pwd.c:
- Use sh_scoped() to obtain the correct value of $OLDPWD.
- Fix a use-after-free bug. Make the 'oldpwd' variable a static
char that points to freeable memory. Each time cd is used, this
variable is freed if it points to a freeable memory address and
isn't also a pointer to shp->pwd.
src/cmd/ksh93/sh/path.c: path_pwd():
- Simplify and add comments.
- Scope $PWD properly.
src/cmd/ksh93/tests/builtins.sh,
src/cmd/ksh93/tests/leaks.sh:
- Backport the ksh2020 regression tests for 'cd -' when $OLDPWD is
set.
- Add test for $OLDPWD and $PWD after subshare.
- Add test for $PWD after 'cd'.
- Add test for possible memory leak.
- Add testing for 'unset' on OLDPWD and PWD.
src/cmd/ksh93/COMPATIBILITY:
- Add compatibility note about changes to $PWD and $OLDPWD.
Co-authored-by: Martijn Dekker <martijn@inlv.org>
This commit adds '/* FALLTHROUGH */' comments to fix many
GCC warnings when compiling with -Wimplicit-fallthrough.
Additionally, the existing fallthrough comments have been
changed for consistency.
src/cmd/ksh93/tests/variables.sh: LC_* error tests:
- Since operating systems validate locale strings differently,
try a few different bad locale strings to find one that makes
setlocale(2) fail, fixing test failures on OpenBSD and Debian.
- Restore warning removed in aed5c6d7, issuing it if none of the
bad locale strings produce a diagnostic.
- Reenable test for diagnostic message disabled in aed5c6d7.
Co-authored-by: Martijn Dekker <martijn@inlv.org>
This bug was originally reported at <https://github.com/att/ast/issues/1467>.
A crash can occur when using the 'b' or 'B' vi mode commands to go back
one word. I was able to reproduce these crashes with 100% consistency on
an OpenBSD virtual machine when ksh is compiled with -D_std_malloc.
Reproducer:
$ set -o vi
$ asdf <ESC> <b or B>
The fix is based on Matthew DeVore's analysis:
> I suspect this is caused by this line:
>> while (vi_isalph(tcur_virt) && tcur_virt >= first_virt) --tcur_virt;
> which is in the b codepath. It checks vi_isalph(tcur_virt) before checking
> if tcur_virt is in range. These two clauses should be reversed. Note that
> line 316 is a similar check for pressing B, and there the tcur_virt value
> is checked first.
src/cmd/ksh93/edit/vi.c:
- Check tcur_virt before using isalph() or isblank() to fix both crashes.
At the start of the backword() while loop this check was performed
twice, so the redundant check has been removed.
src/cmd/ksh93/tests/pty.sh:
- Add a regression test for the b, B, w and W editor commands.
src/cmd/ksh93/bltins/test.c:
- Fix the following compiler warnings from clang:
test.c:554:11: warning: assigning to 'char *' from 'const char []'
discards qualifiers
[-Wincompatible-pointer-types-discards-qualifiers]
e_msg = e_badop;
^ ~~~~~~~
test.c:556:11: warning: assigning to 'char *' from 'const char []'
discards qualifiers
[-Wincompatible-pointer-types-discards-qualifiers]
e_msg = e_unsupported_op;
^ ~~~~~~~~~~~~~~~~
test.c:560:1: warning: control may reach end of non-void function
[-Wreturn-type]
src/cmd/ksh93/tests/builtins.sh:
- Fix regression test by updating error message text.
When test is passed the '=~' operator, it will silently fail with
exit status 1:
$ test foo =~ foo; echo $?
1
This bug is caused by test_binop reaching the 'NOTREACHED' area of
code. The bugfix was adapted from ksh2020:
https://github.com/att/ast/issues/1152
src/cmd/ksh93/bltins/test.c: test_binop():
- Error out with a message suggesting usage of '[[ ... ]]' if '=~'
is passed to the test builtin.
- Special-case TEST_END (']]') as that is not really an operator.
Co-authored-by: Martijn Dekker <martijn@inlv.org>
src/lib/libast/tm/tminit.c:
- Commit 9f43f8d1, in addition to backporting fixes from ksh93v-, also
backported this bug:
$ printf '%(%Z)T' now
PPT # Should be PDT
Reapply the ksh2020 bugfix to fix the %Z time
format again.
src/cmd/ksh93/tests/builtins.sh:
- Add a regression test so this bug (hopefully) isn't backported from
ksh93v- again).
Every so often, a commit's GitHub CI run throws the following
regression test failure:
sigchld.sh[57]: expected '2 background' -- got '3' (DELAY=0.02)
When I re-run the job, the failure usually goes away.
In 712261c8 the DELAY variable was changed from 0.2 to 0.02 to
speed up the first SIGCHLD test. It's possible the GitHub CI
runners are just too slow or too heavily loaded for that.
src/cmd/ksh93/tests/sigchld.sh:
- Restore 0.2 value for 'float DELAY'.
I grepped for #include changes in all the commits and compared
that to the changes in the Mamfiles. I found 7 commits that don't
update the Mamfiles with the appropriate dependencies while
adding #includes, as I only learned how this works after having
worked with this code for some time.
This commit adds the missing Mamfile updates for the
corresponding #include changes in the following commits:
06e721c3, 65d363fd, 70fc1da7, 79d19458, b1a41311, bb4d6a2e,
db71b3ad, and this commit.
Additionally:
src/lib/libast/comp/setlocale.c:
- Change include errno.h to error.h to use EILSEQ fallback if
needed; remove corresponding #ifdef (re: 4dcf5c50, 71bfe028).
src/cmd/ksh93/Mamfile:
- Fix a broken dependency on libast FEATURE/float (re: 72968eae).
We can't use 'prev' for a file that was not mentioned before in
the same Mamfile, we have to use a 'make'...'done' on the first
mention. Add subdependencies matching those in libast/Mamfile.
src/cmd/ksh93/bltins/print.c:
- Rename the unlisted and misleadingly named SHOPT_ECHOE option
(which disables, not enables, 'echo -e') to SHOPT_NOECHOE.
src/cmd/ksh93/SHOPT.sh:
- Add the SHOPT_NOECHOE and SHOPT_TEST_L compile time options to
the list of SHOPT options. Since there is a probe for TEST_L,
set it to probe (empty) by default. NOECHE is off by default.
src/cmd/ksh93/features/options:
- Small bugfix: Allow SHOPT_TEST_L to be manually enabled on
systems that don't support '$(whence -p test) -l /foo'.
- Add a comment describing the SHOPT_MULTIBYTE feature test and
separate it from the SHOPT_DEVFD test.
This bugfix comes from <https://github.com/att/ast/pull/711>.
Eric Scrivner provided the following explanation for the fix:
> Coverity identified an issue with integer truncation in
> `put_enum`. The function was truncating the return values of
> `strcasecmp` and `strcmp` from an `int` to an `unsigned short`
> when assigning them to the local variable `n`. Since either of
> these methods can return a value that is not in the set `{0, 1,
> -1}` the later check if `n == 0` could spuriously evaluate to
> true. For example, in the case where either function returned
> `-65536`.
> The fix is simply to change `n` from an `unsigned short` to an
> `int` to avoid the possibility of truncation. Since the only
> purpose of `n` is the store the return values of these checks,
> this does not have any side effects.
That bit of code supported bash's redundant 'function foo()'
function declaration syntax (with both the 'function' keyword
and the '()') which is a syntax error on ksh, as it should be.
This allows ksh to be compiled with versions of tcc that define
__dso_handle in libtcc1.a, i.e., versions as of this commit:
https://repo.or.cz/tinycc.git/commit/dd60b20c
Older versions of tcc still fail to compile ksh, although now they
fail after reaching the libdll feature test. I'm not sure if fixing
that is feasible since even if I hack out the failing libdll
feature test, ksh fails to link with a '__dso_handle' error.
src/lib/libast/comp/atexit.c,
src/lib/libast/features/lib,
src/lib/libast/vmalloc/vmexit.c:
- From what I've been able to gather the only OSes with support
for on_exit are Linux and SunOS 4. However, on_exit takes two
arguments, so the macro that defines it as taking one argument
is incorrect. Since Solaris (SunOS 5) no longer has this call
and the macro breaks on Linux, the clean fix is to remove it
(atexit(3) is used instead).
src/lib/libast/include/ast.h:
- When compiling with tcc on FreeBSD, pretend to be gcc 2.95.3
instead of gcc 9.3.0. This stops /usr/include/math.h from
activating gcc 3.0+ math compiler builtins that don't exist on
tcc, while still identifying as gcc which is needed to avoid
other FreeBSD system header breakage.
src/cmd/builtin/Mamfile,
src/cmd/builtin/features/pty,
src/lib/libdll/Mamfile,
src/lib/libdll/features/dll:
- tcc forbids combining the -c compiler flag with -l* linker flags.
Use the -lm flag in the iffe feature tests instead of the
Mamfiles. This avoids iffe combining -lm with the -c flag.
src/lib/libast/vmalloc/malloc.c:
- Fix failure to compile with -D_std_malloc.
This patch is from OpenSUSE:
https://build.opensuse.org/package/view_file/shells/ksh/ksh93-malloc-hook.dif
As it turns out tcc needs this change to build ksh with
-D_std_malloc, so it has been applied.
Co-authored-by: Martijn Dekker <martijn@inlv.org>
Resolves: https://github.com/ksh93/ksh/issues/232
src/lib/libast/features/lib,
src/lib/libast/path/pathicase.c:
- FAT32 file systems on Linux don't support FS_CASEFOLD_FL, which
caused globbing to break. Reproducer using a UEFI boot partition:
$ echo /boot/eF*
/boot/eF*
This is fixed by checking for FAT attributes with ioctl, then
checking for FS_CASEFOLD_FL if that fails.
- The check for FS_CASEFOLD_FL didn't work correctly; I still wasn't
able to get --globcasedetect to work on a case-insensitive ext4
folder. Fix that by adding missing parentheses.
Moving the 'err_exit' and 'warning' alias definitions in the
regression tests to one _common file introduced a bug: they are no
longer expanded at compile time when the tests are run with shcomp,
resulting in a 'command not found' (at best) on trying to execute
one. shcomp requires that the alias definitions need to be present
in the file itself. But that means maintaining 50-odd copies again.
I'd rather add a hack to shtests to avoid this.
src/cmd/ksh93/tests/shtests:
- Before running a test with shcomp, physically concatenate _common
and the test script together into a temporary file, minus the '.'
command that includes _common, and compile that with shcomp.
The NOT_USED() macro is already defined in ast.h (which is included
by shell.h) as an alias of NoP(). So it's better to apply the fix
to NoP() so it takes effect for both verrsions, for libast and ksh.
