In private mode, access to previous history is blocked and new history
does not persist and is only available for the duration of the current
session.
This mode can be used when it is not desirable for commandline history
to leak into a session, e.g. via autocomplete or when it is desirable to
test the behavior of fish in the absence of history items without
permanently clearing the history.
I'm sure there are a lot more features that can be incorporated into
private mode, such as restricting access to certain user-specific
configuration files, etc.
This addresses a lot of the concerns raised in #1363 (which was later
changed to track mosh-specific problems). See also #102.
When we discard output because there's been too much, we print a
warning, but subsequent uses of the same buffer still discard.
Now we explicitly reset the flag, so we warn once and everything works
normal after.
Fixes#5267.
For things like
source $undefined
or
source (nooutput)
it was quite annoying that it read from tty.
Instead we now require a "-" as the filename to read from the tty.
This does not apply to reading from stdin if it's redirected, so
something | source
still works.
Fixes#2633.
This adds flags --path and --unpath to builtin set, analogous to
--export and --unexport. These flags change whether a variable is
marked as a path variable.
Universal variables cannot yet be path variables.
This switches quoted expansion like "$foo" to use foo's delimiter instead of
space. The delimiter is space for normal variables and colonf or path variables.
Expansions like "$PATH" will now expand using ':'.
This commit begins to bake in a notion of path-style variables.
Prior to this fix, fish would export arrays as ASCII record separator
delimited, except for a whitelist (PATH, CDPATH, MANPATH). This is
surprising and awkward for other programs to deal with, and there's no way
to get similar behavior for other variables like GOPATH or LD_LIBRARY_PATH.
This commit does the following:
1. Exports all arrays as colon delimited strings, instead of RS.
2. Introduces a notion of "path variable." A path variable will be
"colon-delimited" which means it gets colon-separated in quoted expansion,
and automatically splits on colons. In this commit we only do the exporting
part.
Colons are not escaped in exporting; this is deliberate to support uses
like
`set -x PYTHONPATH "/foo:/bar"`
which ought to work (and already do, we don't want to make a compat break
here).
This reverts commit 3f820f0edf.
While the premise described by @nbuwe is sound in #4505, we are now
apparently relying on this behavior is some places (although
inadvertently as there doesn't seem to be a deliberate acknowledgement
of that anywhere).
Turning off ONLCR causes things like indented multiline commands to not
appear correct at the tty (subsequent lines appear both at column 0 and
again indented).
Per @nbuwe's excellent explanation in #4505, we can save on output
to the tty by maintaining column location after NL by disabling the
ONLCR terminal mode.
Closes#4505.
Adds a new match mode for `string_fuzzy_match_t` that matches against a
case-insensitive subsequence within a string, e.g. `LL` now (partially)
matches against `hello`. This is implemented as a separate mode, given a
lower priority of match than a same-case match (when present).
Note that `fuzzy_match_subsequence_insertions_only` has purposely not
been extended with a case-insensitive version as that would be a)
unlikely to match often, and b) adding a second inefficient fuzzy search
to something that's queried a lot. Perhaps `subsequence_insertions_only`
can simply be changed to be a case-insensitive comparison in the future?
Closes#1196. Affects #3978.
Load fish docs and configuration out of the source and/or build
directories rather from the installed paths when running directly out
of the cmake build directory.
Closes#5255.
Prior to this fix, fish would swallow SIGINT in non-interactive mode. This
meant that scripts could only be Ctrl-C'd if fish was executing an external
command.
Unblock SIGINT in non-interactive mode.
Fixes#5253
Fixes broken macOS build. I'm not sure how the code used to compile
without including `dyld.h` previously, perhaps a different header used
to pull it in?
Retrieves the fully resolved path to the currently executing fish binary
(regardless of PATH). Can be used to ensure that the same fish is
launched again from a script.
`get_executable_path()` moved from fish binary to libfish, also cleaned
up some duplicated (but differing!) definitions of PATH_MAX (which was
used by that function) in the process.
Remove dependency on the Linux compatibility layer's procfs being
installed and mounted when running under FreeBSD by directly querying
the MIB for the path to the running fish executable
(KERN_PROC_PATHNAME). Tested under FreeBSD 11.2-RELEASE.
This switches fish to a "virtual" PWD, where it no longer uses getcwd to
discover its PWD but instead synthesizes it based on normalizing cd against
the $PWD variable.
Both pwd and $PWD contain the virtual path. pwd is taught about -P to
return the physical path, and -L the logical path (which is the default).
Fixes#3350
This new function performs normalization of paths including dropping
/./ segments, and resolving /../ segments, in preparation for switching
fish to a "virtual" PWD.
Mostly resolves#4862, though there remains the lingering question of
whether or not to emit a warning to /dev/tty or stderr when a
non-literal-zero index evaluates to zero.
Coalesces commands with leading (if even possible) and trailing
whitespace into the same item, improving the experience when iterating
over history entries.
Closes#4908.
This allows for marking certain bindings as part of a preset, which allows us to
- only erase those when switching presets
- go back to the preset binding when erasing a user binding
- only show user customization if requested
- make bare bind statements in config.fish work (!!!11elf!!!)
Fixes#5191.
Fixes#3699.
This reverts commit 8c14f0f30f.
This list is not reliable - there are many ways for fish to quit that does not
invoke these functions. It's also not necessary since the history is correctly
saved on exec.
Prior to this change, env_get_pwd_slash() would try to infer the PWD from
getcwd() if $PWD were missing. But this results env_get_pwd_slash() doing
something radically different than $PWD, and also is a lot of code for a
scenario that cannot be reliably reproduced. Just return "/" in this case.
If the replacement in `string replace` is invalid, prior to this fix we would
enter into an infinite loop trying to parse it. Instead report errors correctly.
Fixes#3381
Rather than having tokenizer_error as pointers to objects, switch it back
to just an error code value. This makes reasoning about it easier since
it's immutable values instead of mutable objects, and it avoids allocation
during startup.
At some point the completion code was refactored and in the event where
no explicit function description was passed into `resolve_description()`
it would attempt to use the `desc_func` parameter but pass in the
_remaining_ part of the completion rather than the full text, which
would obviously fail.
e.g. if completing `foo<TAB>`, for function `foobar` it would attempt to
find the description for a function named `bar` instead of `foobar`.
Closes#5206.
This reverts commit 9c63ad3209 until I can
figure out what is causing the assertion and test failures.
It *seems* to be that passing in the correct function name to the
description lookup is causing a previously present error to be realized,
but I can't yet be certain.
At some point the completion code was refactored and in the event where
no explicit function description was passed into `resolve_description()`
it would attempt to use the `desc_func` parameter but pass in the
_remaining_ part of the completion rather than the full text, which
would obviously fail.
e.g. if completing `foo<TAB>`, for function `foobar` it would attempt to
find the description for a function named `bar` instead of `foodbar`.
Closes#5206.
There's been no reproducible case entered for #5080, but the stack trace
indicates the problem is with env_get_pwd_slash() returning an empty
string, which isn't a string that terminates in `/`.
In addition to making the failure case to return the path `./` (which
has the benefit of having the same meaning as $PWD), trying a little bit
harder to retrieve the real PWD by using getcwd(3). While
get_current_dir(3) is documented as relying on PWD, getcwd(3) does not
mention any such caveats, so it's possible that it will work even if
something is breaking PWD.
Just a thought, but it's possible if due to some recursion PWD surpassed
some predetermined value (maybe PATH_MAX) that PWD (on certain platforms
or under certain enivronments) won't be set (hence the code that deals
with ERANGE errors from the getcwd(3) call).
Closes#5080.
As reported in fish-shell/fish-shell#5180, when the USER environment
variable is not set and fish is started, `get_runtime_path()` returns a
blank string. At some point in the past, this was called after
`setup_user()` in env.cpp, but this is no longer the case.
This commit removes the reliance on the $USER environment variable
entirely, and instead uses `getpwuid(geteuid()).pw_name` to retrieve the
current username.
Closes#5180.
`exec` now exhibits the same behavior as `exit` and prompts the user to
confirm their intention to end the current process if there are
background jobs running. Running `exec` again immediately thereafter
will force the exec to go through.
Additionally, background jobs are reaped upon exec to prevent process
leaking (same as `exit`).
Fix#5133 changed builtins to acquire the terminal, but this regressed
caused fish to be stopped when running in background via `sudo fish`.
Fix this by only acquiring the terminal if the terminal was owned by the
builtin's pgroup.
Fixes#5147
- Add support for:
- Jumping to the character before a target.
- Repeating the previous jump (same direction, same precision).
- Repeating the previous jump in the reverse order.
- Enhance vi bindings.
When running a builtin, if we are an interactive shell and stdin is a tty,
then acquire ownership of the terminal via tcgetpgrp() before running the
builtin, and set it back after.
Fixes#4540
Factor the history search fields into a new class.
As a side effect, this shares the deduplication logic, so that token search
no longer returns duplicates.
Fixes#4795
This changes the behavior of builtin math to floating point by default.
If the result of a computation is an integer, then it will be printed as an
integer; otherwise it will be printed as a floating point decimal with up to
'scale' digits past the decimal point (default is 6, matching printf).
Trailing zeros are trimmed. Values are rounded following printf semantics.
Fixes#4478
This reverts commit e2a3dae58b.
This idea failed because ./share was not complete when bliding via cmake;
it misses critical files such as config.fish.
fish tries to be relocatable by looking for directories relative to its
executable. These directories are not found when running fish from
within a cmake build because the etc directory is not present. Stop requiring
this directory to be present since it's not critical for running fish.
Fixes#4825
Don't mmap history files on remote file systems
This merges some changes to history that may help to mitigate the crashes seen in #5088 . These SIGBUS crashes occur when reading a memory mapping whose underlying file was truncated. It's not clear why this should occur more often on NFS (or ever). However memory mapping over NFS is sketchy anyways so this is desirable regardless.
Migrate the mmap() logic into a new class history_file_contents_t which
will serve to encapsulate conditional logic if we choose to use read()
instead of mmap().
This adds a new string command split0, which splits on zero bytes.
split0 has superpowers because its output is not further split on
newlines when used in command substitutions.
separated_buffer_t encapsulates the logic around discarding (which
was previously duplicated between output_stream_t and io_buffer_t),
and will also encapsulate the logic around explicitly separated
output.
If just one of the range ends is negative, this now forces direction away from it.
I.e. if the beginning is negative, we go in reverse.
If the end is negative, we go forwards.
This fixes cases like
$var[2..-1]
if $var only has one element.
While supported by gcc and clang, \e is a gcc-specific extension and not
formally defined in the C or C++ standards.
See [0] for a list of valid escapes.
[0]: https://stackoverflow.com/a/10220539/17027
We've tried numerous approaches to mitigate the race condition between
`posix_spawn` and the `setpgid` call, but unfortunately due to the flags
we pass to `posix_spawn`, it (rarely? never?) results in `vfork()` being
used, which means it is never executed atomically. Since it is executed
out-of-band, we must manually call `setpgid` in case `posix_spawn`
hasn't gotten around to doing that yet, but in the event that it has, an
EACCES error can be returned.
Closes#4884. Closes#4715. See also #4778.
On systems where the terminfo for TERM does not contain a string for
attributes such as enter_underline_mode, etc. fish was crashing with a
fatal error message and a note to email the developers.
These are non-essential text attribute changes and should not trigger
such a failure.
[9/13] Building CXX object CMakeFiles/fishlib.dir/src/builtin_string.cpp.o
../src/builtin_string.cpp:1221:12: warning: mangled name of 'string_transform' will change in C++17 due to non-throwing exception specification in function signature [-Wc++17-compat-mangling]
static int string_transform(parser_t &parser, io_streams_t &streams, int argc, wchar_t **argv, decltype(std::towlower) func) {
^
1 warning generated.
Our completion machinery calls our `__fish_describe_command` function
to describe commands via apropos. Only it trusts the output a bit too
much, so it crashes when any line from that is shorter than the
original string.
Fix this by skipping any string that is shorter than the original,
since it can't be a match anyway.
Also stop doing wcslen so often - std::strings are nice!
Fixes#5014.
There really is no need to
- Timeout just because the _first_ character was a control character
- Timeout because of any control character other than escape
The reason to timeout because the '\e' sequence can appear by itself (signifying
pressing the escape key) and still make
sense - e.g. vi-mode has it bound to a rather important function!
But a \c can't appear by itself, so we can just block.
This allows binding sequences like \cx\ce and inputting them at a
leisurely pace rather than the frantic escape_timeout one.
It should also improve sequences that _include_ escape somewhere else.
E.g. something like a\eb ("a, then alt+b") should now time out for the "\eb" part,
allowing users to bind a\e ("a, then escape") to something else. Why you'd want to do
that, I have no idea. But it's more consistent, and that's nice!
For regex-mode, this should be enough to read NUL-delimited strings to act on, but not
quite patterns and replacements.
Glob-mode requires more work - it uses wcscmp internally, which is unsuitable.
Also the various styles have one function each with barely any
difference - mostly passing the corresponding STYLE argument.
Pack them into one function for escape and one for unescape to save
about 100 lines.
We're now actually handling wchar_t here, so comparing the 0x80 bit
would break for UTF-16, causing ASCII false-positives.
Also simplifies a bit, since we no longer need a second variable.
printf 'a\0b' | string length
used to print "1". Now it prints "3".
Note that this switches to using C++'s std::string::length, which
might give differing results.
The prompt is a fallback that is overridden via a function file
anyway.
Do that with the title as well, so we can use just builtins.
This removes error messages when $fish_function_path is borked.
Introduced by #4849 (add wait for processes by name)
../src/builtin_wait.cpp:23:14: warning: using the result of an assignment as a condition without parentheses [-Wparentheses]
while (j = jobs.next()) {
~~^~~~~~~~~~~~~
../src/builtin_wait.cpp:23:14: note: place parentheses around the assignment to silence this warning
while (j = jobs.next()) {
^
( )
../src/builtin_wait.cpp:23:14: note: use '==' to turn this assignment into an equality comparison
while (j = jobs.next()) {
^
==
1 warning generated.
Turns out the segfaults we've been getting in our tests are because we set $TERM to "dumb".
So we only clear the line if the terminal isn't dumb.
This reverts commit 745a88f2f6.
Fixes#2320.
One key use of process expansion, used in currently-shipped code, is for running a function on
current shell exit.
Restore the use of %self as a valid argument (and add `self`) and document this change.
(faho: Remove bare "self")
This enables users to opt in (or out) of specific features by setting
the fish_features environment variable.
For example `set -U fish_features stderr-nocaret` to opt into removing the
caret redirection.
This partially reverts 5b489ca30f, with
carets acting as redirections unless the stderr-nocaret flag is set.
This flag is off by default but may be enabled on the command line:
fish --features stderr-nocaret
This introduces a new command line option --features which can be used for
enabling or disabling features for a particular fish session.
Examples:
fish --features stderr-nocaret
fish --features 3.0,no-stderr-nocaret
fish --features all
Note that the feature set cannot be changed in an existing session.
This teaches the status command to work with features.
'status features' will show a table listing all known features and whether
they are currently on or off.
`status test-feature` will test an individual feature, setting the exit status to
0 if the feature is on, 1 if off, 2 if unknown.
This introduces a new type features_t that exposes feature flags. The intent
is to allow a deprecation/incremental adoption path. This is not a general
purpose configuration mechanism, but instead allows for compatibility during
the transition as features are added/removed.
Each feature has a user-presentable short name and a short description. Their
values are tracked in a struct features_t.
We start with one feature stderr_nocaret, but it's not hooked up yet.
This was done in share/config.fish, but leads to surprising results if
that isn't read - e.g. because someone just built fish in the git
directory to test it without installing.
It's also not something that is any more or less complicated.
For compatibility, keep it in config.fish as well for the time being.
complete.cpp strips the path from commands before parsing for
completions, meaning that when we called `path_get_path()` against
`cmd`, if `./cmd` were typed in at the command line but `cmd` does not
exist in the PATH, then the command would incorrectly be flagged as not
present and the completions would be skipped.
This is also faster when an absolute/relative path is used for a
command, as we now search with the original path which skips searching
PATH directories unnecessarily.
Found when debugging why completions for `./configure` wouldn't work.
This brings back expansion of `%n` where `n` is a job id, but not as a
general parser syntax. This makes `jobs -p %n` work, which can be used
as part of the job control command chain, i.e.
```
cat &
fg (jobs -p %1)
```
fg/bg/wait can either be wrapped in a function to call `jobs -p` for
`%n` arguments, or they can be updated to take `%n` arguments
themselves.
The order of this list does not need to be strictly maintained any
longer.
Benchmarked with `hyperfine` as follows, where `bench1` is the existing
approach of binary search and `bench2` is the new unordered_set code,
(executed under bash because fish would always return non-zero). The
benchmark code checks each argv to see if it is a builtin keyword (both
return the same result):
```
hyperfine './bench1 $(shuf /usr/share/dict/words)' './bench2 $(shuf /usr/share/dict/words)'
Benchmark #1: ./bench1 $(shuf /usr/share/dict/words)
Time (mean ± σ): 68.4 ms ± 3.0 ms [User: 28.8 ms, System: 38.9 ms]
Range (min … max): 60.4 ms … 75.4 ms
Benchmark #2: ./bench2 $(shuf /usr/share/dict/words)
Time (mean ± σ): 61.4 ms ± 2.3 ms [User: 23.1 ms, System: 39.8 ms]
Range (min … max): 58.1 ms … 67.1 ms
Summary
'./bench2 $(shuf /usr/share/dict/words)' ran
1.11x faster than './bench1 $(shuf /usr/share/dict/words)'
```
Prior to this fix, the fish universal variables file claimed that
changes to it would be overwritten. This no longer true and has not
been true for a long time. Remove that warning.
This switches the universal variables file from a machine-specific
name to the fixed '.config/fish/fish_universal_variables'. The old file
name is migrated if necessary.
Fixes#1912
This removes the caret as a shorthand for redirecting stderr.
Note that stderr may be redirected to a file via 2>/some/path...
and may be redirected with a pipe via 2>|.
Fixes#4394
The previous commit caused the tests to fail since env_remove() was
returning a blanket `!0` when a variable couldn't be unset because it
didn't exist in the first place. This caused the wrong message to be
emitted since the code clashed with a return code for `env_set()`.
Added `ENV_NOT_FOUND` to signify that the variable requested unset
didn't exist in the first place, but _not_ printing the error message
currently so as not to break existing behavior before checking if this
is something we want.
Variables set in if and while conditions are in the enclosing block, not
the if/while statement block. For example:
if set -l var (somecommand) ; end
echo $var
will now work as expected.
Fixes#4820. Fixes#1212.
Currently, there are two possibilities for holes in the background:
- When there are two candidates with the same meaning (a long and a
short option or two candidates with the same description)
- When a candidate does not have a description (meaning the color
won't continue after it)
This changes both so the background just goes on.
In addition, it avoids making the background multiple times.
Fixes#4866.
The official fish documentation makes no mention of how `string split`
treats empty tokens, e.g. splitting 'key1##key2' on '#' or (more
confusingly) splitting '/path' on '/'. With this commit, `string split`
now has an option to exclude zero-length substrings from the resulting
array with a new `--no-empty/-n`. The default behavior of preserving
empty entries is kept so as to avoid breakage.
The two unicode glyphs used to represent missing new lines and redacted
characters for secure entry are both not present in the glyph tables of
the default font under Windows (Consolas and Lucida Console), use an
alternative glyph instead.
The "return" symbol is replaced with a pilcrow (¶) and the "redacted
character" symbol is replaced with a bullet (•). Both of these are
well-defined in almost all fonts as they're very old symbols. This
change only takes place if -DWSL is supplied by the build toolchain.
Note: this means a Windows SSH client connecting to a fish remote
instance on a non-Windows machine will still use the (unavailable)
default glyphs instead.
The newly added `:` command is implemented as a function (to avoid
increasing complexity by making it a builtin), but it is saved to a path
that does not match its filename (since its name is somewhat of a
special character that might cause problems during installation).
Directly probing the `colon` function for autoload causes `:` to be
correctly loaded, so doing just that after function paths are loaded
upon startup.
This is a hack since the CPP code shouldn't really be aware of
individual functions, perhaps there is a better way of doing this.
Fixes an issue introduced in 4414d5c888
where functions loaded from custom directories are not detected as being
valid for purposes of determining whether or not completions should be
called.
Work around #4810 by retrieving localizations at runtime to avoid issues
possibly caused by inserting into the static unordered_map during static
initialization.
Closes#810.
Line continuations (i.e. escaped new lines) now make sense again. With
the smart pipe support (pipes continue on to next line) recently added,
this hack to have continuations ignore comments makes no sense.
This is valid code:
```fish
echo hello |
# comment here
tr -d 'l'
```
this isn't:
```fish
echo hello | \
# comment here
tr -d 'l'
```
Reverts @snnw's 318daaffb2Closes#2928. Closes#2929.
From the discussion in #3802, handling spaces within braces more
gracefully. Leading and trailing whitespace that isn't quoted or escaped
is stripped, whitespace in the middle is preserved. Any whitespace
encountered within expansion tokens is treated as a single space,
similar to how programming languages that don't hard break tokens/quotes
on line endings would.
The value is not electrified or tied and is read-only. It isn't cached
in the get_hostname_identifier() function as the ENV_GLOBAL $hostname
will cache it for its duration.
The behavior of `gethostname` in case of an insufficient buffer is
library and version dependent. Work around this by using a big enough
buffer then truncating the output to our desired max length.
The 0th index of the array was tested inside the loop instead of just
once outside it.
Also explain `input_mapping_is_match` control code behavior and
reasoning and simplify control flow.
Drops the % notation for process expansion. The existing notation was a
mess and expanded jobs, process ids, and process names via dark magic.
With this change, % is no longer a special character and can be used
unescaped with impunity.
The variables %self and %last, referring to fish's own pid and the pid
of the last backgrounded job respectively, have been replaced with $pid
and $last_pid. These are read-only variables, protected against being
redefined by the user.
Author's note: I would have personally preferred $fish_pid instead of
$pid but since we debated changing $version to $fish_version and then
reverted that change (with much acrimony), it makes no sense to break
with that precedent here. Additionally, $fish_last_pid is quite wordy.
Closes#4230. Closes#1202.
When number is infinite, not a number, larger than LONG_MAX or smaller
than LONG_MIN, print a corresponding error and return STATUS_CMD_ERROR.
This should fix the worst of the problems, by at least making them clear.
Fixes#4479.
Fixes#4768.
This allows prompts to react to $COLUMNS by e.g. omitting some parts.
We still fallback to a ">" prompt if that's still not short enough,
but now the user has a way of making a nicer prompt.
Fixes#904.
Fixes#4381.
If the head is not a valid, existent command, do not load and run custom
completion sources. This applies to both the autosuggestion provider and
manual user completions. File-based completions will still be offered.
Supersedes #4782 and #4783. Closes#4783. Closes#4782. Closes#2365.
This promotes "and" and "or" from a type of statement to "job
decorators," as a possible prefix on a job. The point is to rationalize
how they interact with && and ||.
In the new world 'and' and 'or' apply to a entire job conjunction, i.e.
they have "lower precedence." Example:
if [ $age -ge 0 ] && [ $age -le 18 ]
or [ $age -ge 75 ] && [ $age -le 100 ]
echo "Child or senior"
end
This should speed things up on slower PCs given that the vast majority
of shell commands are simple jobs consisting of a single command without
any pipelines, in which case there's no need for a keepalive process at
all. Applies to WSL only.
As a temporary workaround for the behavior described in
Microsoft/WSL#2997 wherein WSL does not correctly assign the spawned
child its own PID as its PGID, explicitly set the PGID for the newly
spawned process.
This now reports "TOO_MANY_ARGS" instead of no error (and triggering
an assertion).
We might want to add a new error type or report the missing operator
before, but this is okay for now.
This turns a bunch of ifs on their heads.
We often see this pattern in te:
```c
if (s->type != SOME_TYPE) {
// error handling
} else {
// normal code
}
```
Only, since we want to return the first error, we do
```c
if (s->type == SOME_TYPE) {
// normal code
} else if (s->type != TOK_ERROR) {
// Add a new error - if it already has type error
// this should already be handled.
}
```
One big issue is the comma operator, that means arity-1 functions can
take an arbitrary number of arguments. E.g.
```fish
math "sin(5,9)"
```
will return the value of sin for _9_, since this is read as "5 COMMA
9".
This enables some limited use of arguments for wrapping completions. The
simplest example is that complete gco -w 'git checkout' now works like
you would want: `gco <tab>` now invokes git's completions with the
`checkout` argument prepended.
Fixes#1976
Previously, in
ls ^a bcd
(with "^" as the cursor), kill-word would delete the "a" and then go
on, remove the space and the "bcd".
With this, it will only kill the "a".
Fixes#4747.
This is part of an effort to improve fish's Unicode handling. This commit
attempts to grapple with the fact that, certain characters (principally
emoji) were considered to have a wcwidth of 1 in Unicode 8, but a width of
2 in Unicode 9.
The system wcwidth() here cannot be trusted; terminal emulators do not
respect it. iTerm2 even allows this to be set in preferences.
This commit introduces a new function is_width_2_in_Uni9_but_1_in_Uni8() to
detect characters of version-ambiguous width. For these characters, it
returns a width guessed based on the value of TERM_PROGRAM and
TERM_VERSION, defaulting to 1. This value can be overridden by setting the
value of a new variable fish_emoji_width (presumably either to 1 or 2).
Fixes#4539, #2652.
`argparse`, `read`, `set`, `status`, `test` and `[` now can't be used
as function names anymore.
This is because (except for `test` and `[`) there is no way to wrap these properly, so any
function called that will be broken anyway.
For `test` (and `[`), there is nothing that can be added and there
have been confused users who created a function that then broke
everything.
Fixes#3000.
Prior to this fix, each redirection type was a separate token_type.
Unify these under a single type TOK_REDIRECT and break the redirection
type out into a new sub-type redirection_type_t.
Prior to this the tokenizer ran "one ahead", where tokenizer_t::next()
would in fact return the last-parsed token. Switch to parsing on demand
instead of running one ahead; this is simpler and prepares for tokenizer
changes.
Turns out the process-exit is only ever used in conjunction with
`%self`. Make that explicit by just adding a new "fish_exit" event,
and deprecate the general process-exit machinery.
Fixes#4700.
The previous attempt to support newlines after pipes changed the lexer to
swallow newlines after encountering a pipe. This has two problems that are
difficult to fix:
1. comments cannot be placed after the pipe
2. fish_indent won't know about the newlines, so it will erase them
Address these problems by removing the lexer behavior, and replacing it
with a new parser symbol "optional_newlines" allowing the newlines to be
reflected directly in the fish grammar.
Prior to this fix, if you attempt to complete from inside a quote and the
completion contained an entity that cannot be represented inside quotes
(i.e. \n \r \t \b), the result would be a broken mess of quotes. Rewrite
the implementation so that it exits the quotes, emits the correct unquoted
escape, and then re-enters the quotes.
Properly escape literal tildes in tab completion results. Currently we
always escape tildes in unquoted arguments; in the future we may escape
only leading tildes.
Fixes#2274
Prior to this fix, autoloads like function and completion autoloads
would check their path variable (like fish_function_path) on every
autoload request. Switch to invalidating it in response to the variable
changing.
This improves time on a microbenchmark:
for i in (seq 50000)
setenv test_env val$i
end
from ~11 seconds to ~6.5 seconds.
The job control functions were a bit messy, in particular
`set_child_group`'s name would imply that all it does is set the child
group, but in reality it used to set the child group (via `setpgid`),
set the job's pgrp if it hasn't been set, and possibly assign control of
the terminal to the newly-created job.
These have been split into separate functions. Now `set_child_group`
does just (and only) that, `maybe_assign_terminal` might assign the
terminal to the new pgrp, and `on_process_created` is used to set the
job properties the first time an external process is created. This might
also speed things up (but probably not noticeably) as there are no more
repeated calls to `getpgrp()` if JOB_CONTROL is not set.
Additionally, this closes#4715 by no longer unconditionally calling
`setpgid` on all new processes, including those created by `posix_spawn`
which does not need this since the child's pgrep is set at in the
arguments to that API call.
This switches function execution from the function's source code to
its stored node and pstree. This means we no longer have to re-parse
the function every time we execute it.
The idea is that we can return the shared pointer directly, avoiding
lots of annoying little getter functions that each need to take locks.
It also helps to pull together the data structures used to initialize
functions versus store them.
This concerns block nodes with redirections, like
begin ... end | grep ...
Prior to this fix, we passed in a pointer to the node. Switch to passing
in the tnode and parsed source ref. This improves type safety and better
aligns with the function-node plans.
Prior to this fix, functions stored a string representation of their
contents. Switch them to storing a parsed source reference and the
tnode of the contents. This is part of an effort to avoid reparsing
a function's contents every time it executes.
Add a fish-specific wrapper around std::mutex that records whether it is
locked in a bool. This is to make ASSERT_IS_LOCKED() simpler (it can just
check the boolean instead of relying on try_lock) which will make Coverity
Scan happier.
Some details: Coverity Scan was complaining about an apparent double-unlock
because it's unaware of the semantics of try_lock(). Specifically fish
asserts that a lock is locked by asserting that try_lock fails; if it
succeeds fish prints an error and then unlocks the lock (so as not to leave
it locked). This unlock is of course correct, but it confused Coverity Scan.
Use wcstring/string instead of a character array. The variable
`term_env` was not being freed before the function exited.
Fixes defect 7520324 in coverity scan.
keepalive processes are typically killed by the main shell process.
However if the main shell exits the keepalive may linger. In WSL
keepalives are used more often, and the lingering keepalives are both
leaks and prevent the tests from finishing.
Have keepalives poll for their parent process ID and exit when it
changes, so they can clean themselves up. The polling frequency can be
low.
Have WSL use a keepalive whenever the first process is external.
This works around the fact that WSL prohibits setting an exited
process as the group leader.
When the pager wants to use the full screen to show many options, it reserves
space at the top to see the command. Previously it pretended the command was a
prompt and engaged the prompt layout mechanism to compute these lines. Instead
let's juts count newlines since escape sequences within commands are very rare.
There were several issues with the way that the include tests for curses.h
were being done that were ultimately causing fish to use the headers from
ncurses but link against curses on platforms that provide an actual
libcurses.so that isn't just a symlink to libncurses.so
In particular, the old code was first testing for curses's cureses.h and then
falling back to libncurses's implementation of the same - but that logic was
reversed when it came to including term.h, in which case it was testing for
the ncurses term.h and falling back to the curses.h header. Long story short,
while cmake will link against libcurses.so if both libcurses.so and
libncurses.so are present (unless CURSES_NEED_NCURSES evaluates to TRUE, but
that makes ncurses a hard requirement), but we were brining in some of the
defines from the ncurses headers, causing SIGSEGV panics when fish ultimately
tried to access variables that weren't exported or were mapped to undefined
areas of memory in the other library.
Additionally it is an error to include termios.h prior to including the plain
Jane curses.h (not ncurses/curses.h), causing errors about unimplemented types
SGTTY/chtype. So far as I can tell, both curses.h and ncurses/curses.h pull in
termios.h themselves so it shouldn't even be necessary to manually include it,
but I have just moved its #include below that of curses.h
This special cases expansion of $history variables, so that slicing
history no longer needs to construct the entire history array. Speedup
is around 100x in my test.
Fixes#4650
Prior to this fix, if the user typed normal characters while the
completion pager was shown, it would begin searching. This feature was
not well liked, so we are going to instead just append the characters as
normal and disable paging. Control-S can be used to toggle the search
field.
Fixes#2249
try_get_child() was taking the address of a reference; clang was thereby
assuming it could not be null and so was dropping the null check. Ensure
we do not dereference a null pointer.
Fixes#4678
This was a symbol that represented either an argument or a redirection.
This was only used as part of argument_or_redirection_list.
It's simpler to just have these types be alternatives in the list type.
Some dotfile users like to add directories to PATH that point at
non-existent directories (because those directories exist on other
machines). Stop warning in that case, unless those directories contain
a colon, in which case it's probably a user error.
Changed cd completion to differentiate between cd autosuggest and cd tab
completion. cd autosuggest will find deepest unique hierarchy and cd tab
completion will not.
Issue #4402
This untangles the CMake versioning issues (I hope) as discussed in #4626.
Note most of the advice found on the Internet about how to inject git
versions into CMake is just wrong.
The behavior we want is to unconditionally run the script
build_tools/git_version_gen.sh at build time (i.e. when you invoke ninja or
make, and not when you invoke cmake, which is build system generation time).
This script is careful to only update the FISH-BUILD-VERSION-FILE if the
contents have changed, to avoid spurious rebuilding dependencies of
FISH-BUILD-VERSION-FILE. Assuming the git version hasn't changed, the script
will run, but not update FISH-BUILD-VERSION-FILE, and therefore
fish_version.o will not have to be rebuilt.
This might normally rebuild more than is necessary even if the timestamp is
not updated, because ninja computes the dependency chain ahead of time. But
Ninja also supports the 'restat' option for just this case, and CMake is rad
and exposes this via BYPRODUCTS. So mark FISH-BUILD-VERSION-FILE as a
byproduct and make the script always update a dummy file
(fish-build-version-witness.txt). Note this is the use case for which
BYPRODUCTS is designed.
We also have fish_version.cpp #include "FISH-BUILD-VERSION-FILE", and do a
semi-silly thing and make FISH-BUILD-VERSION-FILE valid C++ (so there's just
one version file). This means we have to filter out the quotes in other
cases..
This reverts commit 25839b8c36.
This was an attempt to simplify the version generation, but it
computed the version at build sytem generation time rather than
at build time, requiring another run of CMake to update it.
Correctly generate FISH_BUILD_VERSION for use in fish_version.h/cpp and
fish.pc to allow `fish --version` and `echo $version` to work again.
Not needing the same convoluted measures used by Makefile builds to
prevent the regeneration of the fish version file when it hasn't
changed.
Purposely created a new `cmake_git_version_gen.sh` file so that the old
`git_version_gen.sh` remains compatible with the existing Makefile build
script. Same reason why `fish.pc.in` was not modified to use a lowercase
variable name to match the CMAKE variable of the same name.
Closes#4626
A large portion of time was spent constructing strings and passing
them to debug(). Turn debug into a macro so that the strings are only
constructed if they're going to be printed.
This adds a new class arg_iterator_t which encapsulates decisions about
whether to read arguments from stdin or argv. It also migrates the
unread bytes buffer from a static variable to an instance variable.
Profiling with callgrind revealed that about 60% of the time in a `something | string match` call
was actually spent in `string_get_arg_stdin()`,
because it was calling `read` one byte at a time.
This makes it read in chunks similar to builtin read.
This increases performance for `getent hosts | string match -v '0.0.0.0*'` from about 300ms to about 30ms (i.e. 90%).
At that point it's _actually_ quicker than `grep`.
To improve performance even more, we'd have to cut down on str2wcstring.
Fixes#4604.
Prior to this fix, a "bare variable" in math like 'x + 1' would be
looked up in the environment, i.e. equivalent to '$x + 1'. This appears
to have been done for performance. However this breaks the orthogonality
of fish; performance is not a sufficient justification to give math this
level of built-in power, especially because the performance of math is
not a bottleneck. The implementation is also ugly.
Remove this feature so that variables must be prefixed with the dollar
sign and undergo normal variable expansion. Reading 'git grep' output
does not show any uses of this in fish functions or completions.
Also added to changelog.
Fixes#4393
For some reason on Solaris the previous code was refusing to compile
with an error (regarding the declaration of stdout in the opts struct)
error: declaration of ‘__iob’ as array of references
The obvious guess that it had something to do with the name of the
variable in question proved true; renaming it from `stdout` to
`opts.stdout` allows the build to go through.
macOS 10.5 and earlier do not support the convention of returning
a dynamically allocated string, plus this seems like an unnecessary
malloc. Always allocate a buffer for realpath() to write into.
It seems that `parse_cmd_opts` does not correctly handle no arguments,
and so argc was being decremented to -1 causing uninitialized memory
access when argv[0] was dereferenced at a later point.
No longer using `-` to indicate reading to stdout. Use lack of arguments
as stdout indicator. This prevents mixing of variables with stdout
reading and makes it clear that stdout may not be mixed with delimiters
or array mode.
Added an option to read to stdout via `read -`. While it may seem
useless at first blush, it lets you do things like include
mysql -p(read --silent) ...
Without needing to save to a local variable and then echo it back.
Kicks in when `-` is provided as the variable name to read to. This is
in keeping with the de facto syntax for reading/writing from/to
stdin/stdout instead of a file in, e.g., tar, cat, and other standard
unix utilities.
This silences binding errors due to keys not found in the current
termcap config in the default fish bindings.
Closes#4188, #4431, and obviates the original fix for #1155
It was necessary to re-implement builtin_bind as a class in order to
avoid passing around the options array from function to function and
as adding an opts parameter to `get_terminfo_sequence` would require
otps to be passed to all other builtin_bind_ functions so they could, in
turn, pass it to `get_terminfo_sequence`.
On powerpc64 (big-endian platform) one test failed as:
Testing file printf.in ... fail
Output differs for file printf.in. Diff follows:
--- printf.tmp.out 2017-10-02 18:14:17.740000000 -0700
+++ printf.out 2017-10-02 18:11:59.370000000 -0700
@@ -1,5 +1,5 @@
Hello 1 2 3.000000 4.000000 5 6
-a B 0 18446744073709551615
+a B 10 18446744073709551615
It happens due to roughly the following code:
swprintf(..., L"%o", (long long)8);
Here mismatch happens between "%o" (requires 32-bit value)
and 'long long' (requires 64-bit value).
The fix turns it effectively to:
swprintf(..., L"%llo", (long long)8);
as it was previously done for 'x', 'd' and other int-like types.
Makes tests pass on powerpc64.
Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>
The POSIX standard specifies that a buffer should be supplied to
getcwd(), not doing so is undefined (or rather, platform-defined)
behavior. This was causing the getcwd errors on illumos (though not seen
on Solaris 11) reported in #3340Closes#3340
Took care of remaining issues preventing fish from building on Solaris.
Mainly caused by some assumptions that certain defines are POSIX when
they are not (`NAME_MAX`).
Moved `NAME_MAX` defines to common.h - for some reason, it was being
defined in a cpp file (`env_universal_common.cpp`) even though it is used
in multiple source files.
Now compiles on Solaris 11 with GNU Make. Still some warnings because
fish was written with GNU getopt in mind and the Solaris version doesn't
use `const char *` but rather just `char *` for getopt values, but it
builds nevertheless.
Assuming this closes#3340
We had pid_status defined as a pid_t instance, which was fine since on
most platforms pid_t is an alias for int. However, that is not
universally the case and waitpid takes an int *, not a pid_t *.
A completion may have zero length; in this case the length of the
prefix was omitted and the completion was not visible. Correct the
calculation to account for zero-width completions.
Fixes#4424
A recent discussion involving whether `can_be_encoded()` was broken
caused me to notice that we are inconsistent about whether Unicode code
points are specified using `\xXXXX` or `\uXXXX` notation. Which is
harmless but silly and potentially confusing.
This flag was only documented for a few weeks before being renamed
`--show-time` and has been deprecated for a long time. Fish 3.0 is a good
opportunity to remove it.
Instead of treating the search term as a literal string to be matched
treat it as a glob. This allows the user to get a more useful set of
results by using the `*` glob character in the search term.
Partial fix for #3136
* Implement `history search --reverse`
It should be possible to have `history search` output ordered oldest to
newest like nearly every other shell including bash, ksh, zsh, and csh.
We can't make this the default because too many people expect the
current behavior. This simply makes it possible for people to define
their own abbreviations or functions that provide behavior they are
likely used to if they are transitioning to fish from another shell.
This also fixes a bug in the `history` function with respect to how it
handles the `-n` / `--max` flag.
Fixes#4354
* Fix comment for format_history_record()
As discussed in #3805, this patch disables assigning fish to its own
process group at startup. This was trialled in #4349 alongside other
pgrp fixes which introduced additional problems, but this particular fix
seems to be OK.
Fixes#3805 and works around Microsoft/BashOnWindows#1653
* Hoist `for` loop control var to enclosing scope
It should be possible to reference the last value assigned to a `for`
loop control var when the loop terminates. This makes it easier to detect
if we broke out of the loop among other things. This change makes fish
`for` loops behave like most other shells.
Fixes#1935
* Remove redundant line
The type cached_esc_sequences_t caches escape sequences, and is tasked
with finding an escape sequence that prefixes a given string. Before
this fix, it did so by storing the lengths of cached escape sequences,
and searching for substrings of that length. The new implementation
instead stores all cached escape sequences in a sorted vector, and uses
binary search to find the shortest escape sequence that is a prefix of
the input. This is a substantial simplification that also reduces
allocations.
This eliminates the "missing" notion of env_var_t. Instead
env_get returns a maybe_t<env_var_t>, which forces callers to
handle the possibility that the variable is missing.
maybe_t is an implementation of the Maybe/Optional type, allowing
for an optional value to be stored. This will enable a more
principled approach for functions that return values or failure,
such as env_get.
This commit backs out certain optimizations around setting environment
variables, and replaces them with move semantics. env_set accepts a
list, by value, permitting callers to use std::move to transfer
ownership.
Commit f872f25f introduced a freed memory access regression on line 460
of env.cpp, where an environment variable was converted to a temporary
string, the .c_str() address of which was stored while the string
temporary was destroyed.
This commit keeps a reference to the original string lying around so
that the c_str() pointer does not point to freed memory.
Valgrind warns that the sometimes uninitialized sigaction.sa_flags field
is sometimes used when passed to the signal handler.
This patch explicitly zeros out the sigaction.sa_flags field at creation
time.
cherry-picked from krader1961/fish-shell commit b69df4fe72
Fixes#4353 (regression in indexing of history contents) and introduces
new unit tests to catch bad $history indexing in the future.
This implements an LRU cache of recently seen math expressions. When
executing math inside loops and the like this can provide a 33% decrease
in the time to execute the `math` command.
We need our `math` builtin to behave like `bc` with respect to rounding
floating point values to integer to avoid breaking to many existing
uses. So when scale is zero round down to the nearest integer.
Another change for #3157.
The MuParser supports the concept of multiple expressions separated by
commas. This implements support for that so that you can do things like
this:
set results (math '1+1, 4*2, 9^2')
This is the second baby step in resolving #3157. Implement a bare minimum
builtin `math` command. This is solely to ensure that fish can be built
and run in the Travis build environments. This is okay since anyone running
`builtin math` today is already getting an error response.
Also, more work is needed to support bare var references, multiple result
values, etc.
Using a read-only variable like `status` as a for loop control variable
has never worked. But without this change you simply get non-sensical
behavior that leaves you scratching your head in puzzlement. This change
replaces the non-sensical behavior with an explicit error message.
Fixes#4342
Recent changes to switch to unordered sets/maps can cause the order in
which items are returned to be non-deterministic. This change ensures
that the argparse "Mutually exclusive flags" error message to be
deterministic with respect to the order of the interpolated values.
Make setting fish vars more efficient by avoiding creating a
wcstring_list_t for the case where we're setting one value. For the case
where we're passing a list of values swap it with the list in the var
rather than copying it. This makes the benchmark in #4200 approximately
6% faster.
Since we are including XXHash32/64 anyway for the wchar_t* hashing,
we might as well use it.
Use arch-specific hash size and xxhash for all wcstring hashing
Instead of using XXHash64 for all platforms, use the 32-bit version
when running on 32-bit platforms where XXHash64 is significantly slower
than XXHash32 (and the additional precision will not be used).
Additionally, manually specify wcstring_hash as hashing method for
non-const wcstring unordered_set/map instances (the const varieties
don't have an in-library hash and so already use our xxhash-based
specialization when calling std::hash<const wcstring>).
commit 50f414a45d58fcab664ff662dd27befcfa0fdd95
Author: Mahmoud Al-Qudsi <mqudsi@neosmart.net>
Date: Sat Aug 19 13:43:35 2017 -0500
Converted file_id_t set to unordered_set with custom hash
commit 83ef2dd7cc1bc3e4fdf0b2d3546d6811326cc3c9
Author: Mahmoud Al-Qudsi <mqudsi@neosmart.net>
Date: Sat Aug 19 13:43:14 2017 -0500
Converted remaining set<wcstring> to unordered_set<wcstring>
commit 053da88f933f27505b3cf4810402e2a2be070203
Author: Mahmoud Al-Qudsi <mqudsi@neosmart.net>
Date: Sat Aug 19 13:29:21 2017 -0500
Switched function sets to unordered_set
commit d469742a14ac99599022a9258cda8255178826b5
Author: Mahmoud Al-Qudsi <mqudsi@neosmart.net>
Date: Sat Aug 19 13:21:32 2017 -0500
Converted list of modified variables to an unordered set
commit 5c06f866beeafb23878b1a932c7cd2558412c283
Author: Mahmoud Al-Qudsi <mqudsi@neosmart.net>
Date: Sat Aug 19 13:15:20 2017 -0500
Convert const_string_set_t to std::unordered_set
As it is a readonly-list of raw character pointer strings (not
wcstring), this necessitated the addition of a hashing function since
the C++ standard library does not come with a char pointer hash
function.
To that end, a zlib-licensed [0] port of the excellent, lightweight
XXHash family of 32- and 64-bit hashing algorithms in the form of a C++
header-only include library has been included. XXHash32/64 is pretty
much universally the fastest hashing library for general purpose
applications, and has been thoroughly vetted and is used in countless
open source projects. The single-header version of this library makes it
a lot simpler to include in the fish project, and the license
compatibility with fish' GPLv2 and the zero-lib nature should make it an
easy decision.
std::unordered_set brings a massive speedup as compared to the default
std::set, and the further use of the fast XXHash library to provide the
string hashing should make all forms of string lookups in fish
significantly faster (to a user-noticeable extent).
0: http://create.stephan-brumme.com/about.html
commit 30d7710be8f0c23a4d42f7e713fcb7850f99036e
Author: Mahmoud Al-Qudsi <mqudsi@neosmart.net>
Date: Sat Aug 19 12:29:39 2017 -0500
Using std::unordered_set for completions backing store
While the completions shown to the user are sorted, their storage in
memory does not need to be since they are re-sorted before they are
shown in completions.cpp.
commit 695e83331d7a60ba188e57f6ea0d9b6da54860c6
Author: Mahmoud Al-Qudsi <mqudsi@neosmart.net>
Date: Sat Aug 19 12:06:53 2017 -0500
Updated is_loading to use unordered_set
No longer using RAII wrappers around pthread_mutex_t and pthread_cond_t
in favor of the C++11 std::mutex, std::recursive_mutex, and
std::condition_variable data types.
The `react_to_variable_change()` function is called whenever a fish var
is set. Even as a consequence of statements like `for x in a b c`. It is
therefore critical that that function be as fast as possible. Especially
when setting the var doesn't have any side-effects which is true something
like 99.9999% of the time.
This change reduces the overhead of `react_to_variable_change()` to
unmeasurable levels. Making the synthetic benchmark in issue #4341
36% faster.
Fixes#4341
Internally fish should store vars as a vector of elements. The current
flat string representation is a holdover from when the code was written
in C.
Fixes#4200
A semi-empty var is one with a single empty string element. The
`env_var_t::empty()` method returns true for such vars but we want
`set --show` to report that it has a single empty element.
Newer versions of GCC and Clang are not satisfied by a cast to void,
this fix is adapted from glibc's solution.
New wrapper function ignore_result should be used when a function with
explicit _unused_attribute_ wrapper is called whose result will not be
handled.
Make the `env_var_t::missing_var()` object a singleton rather than a
dynamically constructed object. This requires some discipline in its use
since C++ doesn't directly support immutable objects. But it is slightly
more efficient and helps identify code that incorrectly mutates `env_var_t`
objects that should not be modified.
It's bugged me forever that the scope is the second arg to `env_get()`
but not `env_set()`. And since I'll be introducing some helper functions
that wrap `env_set()` now is a good time to change the order of its
arguments.
My previous change to eliminate `class var_entry_t` caused me to notice
that `env_get()` turned a set but empty var into a missing var. Which
is wrong. Fixing that brought to light several other pieces of code that
were wrong as a consequence of the aforementioned bug.
Another step to fixing issue #4200.
* Fix clearing abandoned line with VTE
With VTE-based terminals, resizing currently causes multi-line prompts
to go weird.
This changes the sequence we use to clear the line to one suggested by
a VTE
developer (https://bugzilla.gnome.org/show_bug.cgi?id=763390#c4).
It changes nothing in konsole 17.04.3 and urxvt 9.22, but they already
work.
Note that this does not fix the case where output did not end in a
newline, but that doesn't seem to be up to us. Also, it only affects
those lines.
Fixes#2320.
* Use terminfo definition instead of hardcoding
Thanks to @ixjlyons.
Doing `set -U var` when a global named `var` exists can result in
confusing behavior. Try to limit the confusion by improving the warning
we write. Also, only write the warning if interactive.
Fixes#4267
The process_t pointer sent to setup_child_process can actually be 0
without it being failure, as that is what fish sends when `exec` is run
(in the case of INTERNAL_EXEC).
This was causing exec to fail.
There is no more race condition between parent and child with
regards to setting the process groups. Each child sets it for themselves
and then blocks indefinitely until the parent does what it needs to for
them (having waited for them to set their process groups). They are not
SIGCONT'd until the next process in the chain (if any) starts so that
that process can join their process group and open the pipes.
In the last commit, we introduced an indiscriminate if !EXTERNAL check
that unblocks a previously SIGSTOP'd command (if any) to allow the main
loop in exec_job to read from it without deadlocking (since builtins and
functions read directly from input as an optimization, sometimes).
Now only unblocking where a fork will not happen to ensure that if a
builtin ends up forking, that fork'd process is guaranteed to be able to
join the previous process' process group and access its output pipes.
Setting the process group in a fork/exec scenario is a well-documented
race condition in pretty much any job control mechanism [0] [1]. The
Wikipedia article contradicts the glibc article and suggests that the
best approach is for the parent to wait for the child to become the
process group leader, while the glibc article suggests that both should
make it so (which is what fish did previously). However, I'm running
into cases where tcsetpgrp is causing an EPERM error, which it isn't
documented to do except if the session id for the calling process
differs from that of the target process group (which is never the case
in fish since they are all part of the same session), which should cause
a _different_ error (SIGTTOU to be sent to all members of the calling
process' group).
In all cases, this is easily remedied by checking if the process group
in question is already in control of the terimnal. There's still the
off-chance that in the time between we check that and the time that the
command completes that situation may have changed, but the parent
process is supposed to ignore the result of this call if it errors out.
[0]: https://en.wikipedia.org/wiki/Process_group
[1]: https://www.gnu.org/software/libc/manual/html_node/Launching-Jobs.html
We were having child processes SIGSTOP themselves immediately after
setting their process group and before launching their intended targets,
but they were not necessarily stopped by the time the next command was
being executed (so the opposite of the original race condition where
they might have finished executing by the time the next command came
around), and as a result when we sent them SIGCONT, that could never
reach. Now using waitpid to synchronize the SIGSTOP/SIGCONT between the
two.
If we had a good, unnamed inter-process event/semaphore, we could use
that to have a child process conditionally stop itself if the next
command in the job chain hadn't yet been started / setup, but this is
probably a lot more straightforward and less-confusing, which isn't a
bad thing.
Additionally, there was a bug caused by the fact that the main exec_job
loop actually blocks to read from previous commands in the job if the
current command is a built-in that doesn't need to fork.
With this waitpid code, I was able to finally add the SIGSTOP code to
all the fork'd processes in the main exec_job loop without introducing
deadlocks; it turns out that they should be treated just like the main
EXTERNAL fork, but they tend to execute faster causing the same deadlock
described above to occur more readily.
The only thing I'm not sure about is whether we should execute
unblock_pid undconditionally for all !EXTERNAL commands. It makes more
sense to *only* do that if a blocking read were about to be done in the
main loop, otherwise the original race condition could still appear
(though it is probably mitigated by whatever duration the SIGSTOP lasted
for, even if it is SIGCONT'd before the next command tries to join the
process group).