This counter is documented as follows:
"The current depth for subshells and command substitution."
But before this commit, the actual behaviour was that the counter
was reset to zero whenever a subshell forked for any reason: a
pipe, background job, running 'ulimit', redirecting stdout in a
command substitution, and more. This behaviour was:
1. Not consistent with the documentation. Non-forked (a.k.a.
virtual) subshells are an internal implementation detail which
scripts should not have to be concerned with. The manual page
doesn't mention them at all.
2. Inherently broken. Since a subshell may fork for any number of
reasons, even mid-run, and those reasons may change with
bugfixes and further development, scripts have never actually
been able to rely on the value of ${.sh.subshell}.
So, this commit fixes the counter to count the levels of all
subshells, both virtual and forked.
src/cmd/ksh93/sh/xec.c: _sh_fork():
- Increase ${.sh.subshell} whenever we fork.
src/cmd/ksh93/sh/subshell.c:
- sh_subfork():
* Fix comment to properly explain what it does. It doesn't
"create" a subshell, it forks off an existing virtual subshell.
* Don't zero ${.sh.subshell}. Instead, since sh_fork() increases
it but we're forking an existing subshell, undo the increase.
- sh_subshell():
* Remove 'int16_t subshell' variable. It was unnecessary and
mostly unused. It was also the wrong type: it was assigned the
value from shp->subshell which is of type short.
* Increase and decrease the level of virtual subshells and
${.sh.subshell} independently.
src/cmd/ksh93/tests/variables.sh:
- Add regression tests for ${.sh.subshell} in virtual and forked
subshells of several kinds: comsub, parentheses, pipe, bg job.
- Undo wrong error test count fix from 04b4aef0.
(cherry picked from commit a0e0e29e7e0dbf21e4b3958ae02bde6665fb2696)
The sfputr() function (put out a null-terminated string) contained
a use of memccpy() that was invalid and could cause crashes,
because the buffer it was copying into could overlap or even be
identical with the buffer being copied from.
Among (probably) other things, this commit fixes a crash in 'print
-v' (print a compound variable structure) on macOS, that caused the
comvar.sh and comvario.sh regression tests to fail spectacularly.
Now they pass.
Issue discovered and fixed by Kurtis Rader in the abandoned ksh2020
branch; this commit backports the fix. He wrote:
| #if _lib_memccpy && !__ia64 /* these guys may never get it right */
|
| The problem is that assertion is wrong. It implies that the libc
| implementation of memccpy() on IA64 is broken. Which is
| incorrect. The problem is the AST sfputr() function is depending
| on what is explicitly undefined behavior in the face of
| overlapping source and destination buffers.
| [...] Using memccpy() simply complicates the code and is unlikely
| to be measurably, let alone noticeably, faster.
Further discussion/analysis: https://github.com/att/ast/issues/78
src/lib/libast/sfio/sfputr.c:
- Remove memccpy use. Always use the manual copying loop.
(cherry picked from commit fbe3c83335256cb714a4aa21f555083c9f1d71d8)
This fixes (or at least works around) a bug that caused special
variables such as PATH, LANG, LC_ALL, LINENO, etc. to lose their
effect after being unset in a subshell.
For example:
(unset PATH; PATH=/dev/null; ls); : wrongly ran 'ls'
(unset LC_ALL; LC_ALL=badlocale); : failed to print a diagnostic
This is yet another problem with non-forking/virtual subshells. If
you forced the subshell to fork (one way of doing this is using the
'ulimit' builtin, e.g. ulimit -t unlimited) before unsetting the
special variable, the problem vanished.
I've tried to localise the problem. I suspect the sh_assignok()
function, which is called from unall(), is to blame. This function
is supposed to make a copy of a variable node in the virtual
subshell's variable tree. Apparently, it fails to copy the
associated permanent discipline function settings (stored in the
np->nvfun->disc pointer) that gave these variables their special
effect, and which survive unset. However, my attempts to fix that
have been unsuccessful. If anyone can figure out a fix, please send
a patch/pull request!
Data point: This bug existed in 93u 2011-02-08, but did not yet
exist in M-1993-12-28-s+. So it is a regression.
Meanwhile, pending a proper fix, this commit adds a safe
workaround: it forces a non-forked subshell to fork before
unsetting such a special variable.
src/cmd/ksh93/bltins/typeset.c: unall():
- If we're in a non-forked, non-${ ...; } subshell, then before
unsetting any variable, check for variables with internal
trap/discipline functions, and call sh_subfork() if any are
found. To avoid crashing, this must be done before calling
sh_pushcontext(), so we need to loop through the args separately.
src/cmd/ksh93/tests/variables.sh:
- Remove the 'ulimit' that forced the fork; we do this in C now.
(cherry picked from commit 21b1a67156582e3cbd36936f4af908bb45211a4b)
The $! special parameter was not set if a background job
(somecommand &) or co-process (somecommand |&) was launched as the
only command within a braces block with an attached redirection,
for example:
{
somecommand &
} >&2
With the bug, $! was unchanged; now it contains the PID of
somecommand.
Ref.: https://github.com/att/ast/issues/1357
src/cmd/ksh93/sh/parse.c: item():
- When processing redirections following a compound command, always
create a parent node with the TSETIO (I/O redirection) token.
Before this commit, if the last command was of type TFORK (and
the last command only tested as TFORK if the bg job or coprocess
was the only command in a braces block, because the ksh parser
optimises away the braces in that case), then the parent node was
created with the TFORK token instead.
I have no idea what David Korn's intention was with that, but
this is clearly very wrong. Creating another TFORK node when
parsing the redirection caused sh_exec() in sh/xec.c to execute
the redirection in an extra forked, non-background subshell.
Since redirections are executed before anything else, this
subshell is what then launched the background job between the
braces, so $! (a.k.a. shp->bckpid) was updated in that subshell
only, and never in the main shell. The extra subshell also
prevented the background job from being noticed by job control
on interactive shells.
So, the fix is simply to remove the broken test for TFORK.
src/cmd/ksh93/tests/variables.sh:
- Add regression tests for a bg job and a co-process as the only
command within a braces block with attached redirection.
(cherry picked from commit ffe5df30e69f7b596941a98498014d8e838861f2)
Fix a bug in autoloading functions. Directories in the path search
list which should be skipped (e.g. because they don't exist) did
not interact correctly with autoloaded functions, so that a
function to autoload was not always found.
Details:
https://github.com/att/ast/issues/1454
Fix backported (and cleaned up) from:
https://github.com/att/ast/commit/3bc58164
src/cmd/ksh93/sh/path.c:
- path_opentype(): Fix the path search loop so that entries marked
with PATH_SKIP are handled correctly.
src/cmd/ksh93/tests/functions.sh:
- Add regression test verifying an autoloaded function with a PATH
that triggered the bug.
The bug in path_opentype() fixed by this commit may affect other
scenarios but we know it affects autoloaded functions. Hence the
test for that scenario.
(cherry picked from commit a27903165775309f4f032de5d42ec1785f14cfbc)
Fix BUG_TESTERR1A: POSIX non-compliance of 'test'/'[' exit status
on error. The command now returns status 2 instead of 1 when given
an invalid number or arithmetic expression, e.g.: [ 123 -eq 123x ]
The problem was that the test builtin (b_test()) calls the generic
arithmetic evaluation subsystem (sh/arith.c, sh/streval.c) which
has no awareness of the test builtin. A simple solution would be to
always make the arithmetic subsystem use an exit status > 1 for
arithmetic errors, but globally changing this may cause backwards
compatibility issues. So it's best to change the behaviour of the
'test' builtin only. This requires the arithmetic subsystem to be
aware of whether it was called from the 'test' builtin or not. To
that end, this commit adds a global flag and overrides the
ERROR_exit macro where needed.
src/cmd/ksh93/include/defs.h,
src/cmd/ksh93/sh/defs.c:
- Declare and initialise a global sh_in_test_builtin flag.
- Declare internal function for ERROR_exit override in test.c.
src/cmd/ksh93/bltins/test.c:
- Add override for ERROR_exit macro using a function that checks if
the exit status is at least 2 if the error occurred while running
the test builtin.
- b_test(): Set sh_in_test_builtin flag while running test builtin.
src/cmd/ksh93/sh/arith.c,
src/cmd/ksh93/sh/streval.c:
- Override ERROR_exit macro using function from test.c.
src/cmd/ksh93/tests/bracket.sh:
- Add regression test verifying status > 1 on arith error in test.
(cherry picked from commit 5eeae5eb9fd5ed961a5096764ad11ab870a223a9)
Aliases can now be correctly unset within subshell environments
(such as ( ... ), $(command substitutions), etc), as well as
non-subshell "shared" command substitutions (${ ...; }). Before,
attempts to unset aliases within these were silently ignored.
Prior discussion: https://github.com/att/ast/issues/108
Subshell alias trees are only referenced in a few places in the
code, *and* have always been broken, so this commit gets rid of the
whole notion of a subshell alias tree. Instead, there is now just
one flat alias tree, and subshells fork into a separate process
when aliases are set or unset within them. It is not really
conceivable that this could be a performance-sensitive operation,
or even a common one, so this is a clean fix with no downside.
src/cmd/ksh93/include/defs.h:
- Remove sh_subaliastree() definition.
src/cmd/ksh93/sh/subshell.c:
- Remove salias element (pointer to subshell alias tree) from
subshell struct.
- Remove sh_subaliastree() function.
- sh_subshell(): Remove alias subshell tree cleanup.
src/cmd/ksh93/bltins/typeset.c:
- b_alias(): If in subshell, fork before setting alias.
- b_unalias(): If in subshell, fork before unsetting alias.
- unall(): Remove sh_subaliastree() call.
src/cmd/ksh93/sh/name.c:
- nv_open(): Remove sh_subaliastree() call.
src/cmd/ksh93/tests/subshell.sh:
- Add regression tests for unsetting or redefining aliases within
subshells.
(cherry picked from commit 12a15605b9521a2564a6e657905705a060e89095)
Functions can now be correctly redefined and unset in subshell
environments (such as ( ... ), $(command substitutions), etc).
Before this fix, attempts to do this were silently ignored (!!!),
causing the wrong code (i.e.: the function by the same name from
the parent shell environment) to be executed.
Redefining and unsetting functions within "shared" command
substitutions of the form '${ ...; }' is also fixed.
Prior discussion: https://github.com/att/ast/issues/73
src/cmd/ksh93/sh/parse.c:
- A fix from George Koelher (URL above). He writes:
| The parser can set t->comnamp to the wrong function.
| Suppose that the shell has executed
| foo() { echo WRONG; }
| and is now parsing
| (foo() { echo ok; } && foo)
| The parser was setting t->comnamp to the wrong foo. [This
| fix] doesn't set t->comnamp unless it was a builtin. Now the
| subshell can't call t->comnamp, so it looks for foo and finds
| the ok foo in the subshell's function tree.
src/cmd/ksh93/bltins/typeset.c:
- Unsetting functions in a virtual/non-forked subshell still
doesn't work: nv_open() fails to find the function. To work
around this problem, make 'unset -f' fork the subshell into its
own process with sh_subfork().
- The workaround exposed another bug: if we unset a function in a
subshell tree that overrode a function by the same name in the
main shell, then nv_delete() exposes the function from the main
shell scope. Since 'unset -f' now always forks a subshell, the
fix is to simply walk though troot's parent views and delete any
such zombie functions as well. (Without this, the 4 'more fun'
tests in tests/subshell.sh fail.)
src/cmd/ksh93/sh/subshell.c: sh_subfuntree():
- Fix function (re)definitions and unsetting in "shared" command
substitutions of the form '${ commandlist; }' (i.e.: if
sp->shp->subshare is true). Though internally this is a weird
form of virtual subshell, the manual page says it does not
execute in a subshell (meaning, all changes must survive it), so
a subshell function tree must not be created for these.
src/cmd/ksh93/tests/subshell.sh:
- Add regression tests related to these bugfixes. Test unsetting
and redefining a function in all three forms of virtual subshell.
(cherry picked from commit dde387825ab1bbd9f2eafc5dc38d5fd0bf9c3652)
Certain environment variables were interpreted as arithmetic
expressions on startup, leading to code injection.
Ref.:
https://bugzilla.redhat.com/show_bug.cgi?id=1757324c7de8b6412
(cherry picked from commit ee6b001d0611ad2e00b6da2c2b42051995c0a678)
The issue with truncating files was caused by out-of-sync streams.
Details and discussion: https://github.com/att/ast/issues/61
src/cmd/ksh93/sh/io.c: sh_iorestore():
- To be safe, sync all streams before restoring file descriptors.
src/cmd/ksh93/tests/io.sh:
- Add two regression tests for truncating files with this
combination of redirections.
- The second test, which invokes a -c script, is disabled for now
as this triggers another corner case bug involving the SH_NOFORK
optimisaton for -c scripts. That fix is for another commit.
(cherry picked from commit 18fb64840365c2ff4608188e5487bd79d08f67d1)
This fixes three related bugs:
1. Expansions like ${var+set} remained static when used within a
'for', 'while' or 'until' loop; the expansions din't change
along with the state of the variable, so they could not be used
to check whether a variable is set within a loop if the state of
that variable changed in the course of the loop. (BUG_ISSETLOOP)
2. ${IFS+s} always yielded 's', and [[ -v IFS ]] always yielded
true, even if IFS is unset. (BUG_IFSISSET)
3. IFS was incorrectly exempt from '-u' ('-o nounset') checks.
src/cmd/ksh93/sh/macro.c: varsub():
- When getting a node pointer (np) to the parameter to test,
special-case IFS by checking if it has a value and not setting
the pointer if not. The node to IFS always exists, even after
'unset -v IFS', so before this fix it always followed the code
path for a parameter that is set. This fixes BUG_IFSISSET for
${IFS+s} and also fixes set -u (-o nounset) with IFS.
- Before using the 'nv_isnull' macro to check if a regular variable
is set, call nv_optimize() if needed. This fixes BUG_ISSETLOOP.
Idea from Kurtis Rader: https://github.com/att/ast/issues/1090
Of course this only works if SHOPT_OPTIMIZE==1 (the default),
but if not, then this bug is not triggered in the first place.
- Add some comments for future reference.
src/cmd/ksh93/bltins/test.c: test_unop():
- Fix BUG_IFSISSET for [[ -v IFS ]]. The nv_optimize() method
doesn't seem to have any effect here, so the only way that I can
figure out is to special-case IFS, nv_getval()'ing it to check if
IFS has a value in the current scope.
src/cmd/ksh93/tests/variables.sh:
- Add regression tests for checking if a varariable is set within a
loop, within and outside a function with that variable made local
(to check if the scope is honoured). Repeat these tests for a
regular variable and for IFS, for ${foo+set} and [[ -v foo ]].
(cherry picked from commit a2cf79cb98fa3e47eca85d9049d1d831636c9b16)
'command -p' was broken for non-interactive shells as the variable
used to store the default system PATH, std_path, was not
initialised correctly. For instance:
$ ksh -c 'command -p ls'
ksh: ls: not found
This fix by Siteshwar Vashisht is backported from ksh2020.
Ref.:
https://github.com/att/ast/issues/426https://github.com/att/ast/pull/448
src/cmd/ksh93/sh/path.c:
- Correctly initialise std_path (the default PATH) when ksh is
started as a non-interactive shell.
src/cmd/ksh93/sh.1:
- Fix vague explanation of 'command -p'.
src/cmd/ksh93/tests/path.sh:
- Add regression test.
(cherry picked from commit a76439d60b70c18cf44d84c1962fcd8df84c947c)
Standard output (FD 1) tested as being on a terminal within a
command substitution, which makes no sense as the command
substitution is supposed to be catching standard output.
ksh -c 'v=$(echo begincomsub
[ -t 1 ] && echo oops
echo endcomsub)
echo "$v"'
This should not output "oops".
This is one of the many bugs with ksh93 virtual (non-forked)
subshells. On the abandoned Vashist/Rader ksh2020 branch, this bug
was fixed by changing quite a lot of code, which introduced and/or
exposed another bug:
https://github.com/att/ast/issues/1079https://github.com/att/ast/commit/8e1e405ehttps://github.com/att/ast/issues/1088
That issue was unresolved when the ksh2020 branch was abandoned.
The safer and more conservative fix is simply forcing the subshell
to fork if we're in a non-forked command substitution and testing
'-t 1'. It is hard to imagine a situation where this would cause a
noticable performance hit.
Note that this fix does not affect ksh93-specific "shared"
non-subshell ${ command substitutions; } which are executed in the
main shell environment, so that variables survive, etcetera.
'test -t 1' continues to wrongly return true there, but command
substitutions of that form cannot be forked because that would
defeat their purpose.
src/cmd/ksh93/bltins/test.c:
- Fix 'test -t 1', '[ -t 1 ]' and '[[ -t 1 ]]' by forking the
current subshell if it is a virtual/non-forked subshell
(shp->subshell), and a command substitution (shp->comsub), but
NOT a "shared" ${ command substitution; } (!shp->subshare).
src/cmd/ksh93/tests/bracket.sh:
- Add two regression tests for this issue, which were adapted from
the Vashist/Rader ksh2020 branch.
NEWS, src/cmd/ksh93/include/version.h:
- Update.
(cherry picked from commit b8ef05e457ead65b83417699b8dd8632f855e2fa)
This new wrapper script sets up the correct environment for running
the ksh93 regression test suite. It allows running the tests
without AST nmake, which is not maintained in this repository.
An alternative ksh to test may be passed in the $KSH env var.
bin/shtests:
- Added. Sets up environment before passing control to
'src/cmd/ksh93/tests/shtest'. Passes on any options given.
NEWS:
- Updated.
(cherry picked from commit 14ced94ed83991687c645a09bd2e45a5c2ffe8dc)
This allows scripts to check for a nonzero exit status on the
'print', 'printf' and 'echo' builtins and prevent possible infinite
loops if SIGPIPE is ignored.
sfsync() was already returning a negative value on I/O error, so
all we need to do is add a check. The stream buffer will need to be
filled before an I/O error can be detected, but this is the same on
other shells. See manual page: src/lib/libast/man/sfio.3
Ref.: https://github.com/att/ast/issues/1093https://github.com/att/ast/pull/1363
src/cmd/ksh93/bltins/print.c: b_print():
- Make sure an error result from sfsync() is reflected in the
output builtin's exit status (exitval).
- Write an I/O error message (e_io) if the exit status is nonzero.
src/cmd/ksh93/data/msg.c, src/cmd/ksh93/include/io.h:
- Add the e_io[] error message.
src/cmd/ksh93/tests/builtins.sh:
- Add I/O error regression test, checking for the correct error
message and exit status. All three output builtins use the same
b_print() function so we only need to test one.
src/cmd/ksh93/tests/basic.sh,
src/cmd/ksh93/tests/coprocess.sh:
- Redirect stderr on a few 'print' commands to /dev/null; these
now issue an expected I/O error. This does not cause failures.
NEWS, TODO:
- Update.
(cherry picked from commit 9011fa933552e483dab460f7dd1593d64e059d94)
This fix is backported from the Vashisht/Rader ksh2020 branch.
src/cmd/ksh93/bltins/print.c:
- Fix syncing history when print -s -f is used. For example, the
following now correctly adds a 'cd' command to the history:
print -s -f 'cd -- %q\n' "$PWD"
Ref.: https://github.com/att/ast/issues/425https://github.com/att/ast/pull/442
src/cmd/ksh93/include/version.h:
- Version date bump.
(cherry picked from commit 46ba7ecdc5c14cd73f6cb24b16c50bdc331a000e)
ksh used to redirect standard output by default when no file
descriptor was specified with the rarely used '<>' reading/writing
redirection operator. It now redirects standard input by default,
as POSIX specifies and as all other POSIX shells do. To redirect
standard output for reading and writing, you now need '1<>'.
Ref.: https://github.com/att/ast/issues/75http://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html#tag_18_07_07
(cherry picked from commit 29afc16c47824fc79ed092ae7704c525b1db6a0a)
This fixes an undocumented 'case' pattern matching misbehaviour
(labelled BUG_CASELIT in modernish) that goes back to the original
Bourne shell, but wasn't discovered until 2018.
If a pattern doesn't match as a pattern, it's tried again as a
literal string. This breaks common validation use cases, such as:
n='[0-9]'
case $n in
( [0-9] ) echo "$n is a number" ;;
esac
would output "[0-9] is a number" as the literal string fallback
matches the pattern. As this misbehaviour was never documented
anywhere (not for Bourne, ksh88, or ksh93), and it was never
replicated in other shells (not even in ksh88 clones pdksh and
mksh), it is unlikely any scripts rely on it.
Of course, a literal string fallback, should it be needed, is
trivial to implement correctly without this breakage:
case $n in
( [0-9] | "[0-9]") echo "$n is a number or the number pattern" ;;
esac
src/cmd/ksh93/sh/xec.c:
- Remove trim_eq() function responsible for implementing the
misbehaviour described above.
NEWS:
- Added. Document this bugfix.
Ref.:
- The problem: thread starting at
https://www.mail-archive.com/austin-group-l@opengroup.org/msg02127.html
- The solution, thanks to George Koehler: comments/commits in
https://github.com/att/ast/issues/476
- Modernish BUG_CASELIT bug test & documentation:
https://github.com/modernish/modernish/commit/b2024ae3
(cherry picked from commit 8d6c8ce69884767a160c1e20049e77bdd849c248
with some extra edits to NEWS to upate the info for this reboot)
