When the user clicks somewhere in the prompt, kitty asks the shell
to move the cursor there (since there is not much else to do).
This is currently implemented by sending an array of
forward-char-passive commands. This has problems, for example it
is really slow on large command lines (probably because we repaint
everytime).
Implement kitty's `click_events=1` flag to set the
position directly. To convert from terminal-coordinates
to fish-coordinates, query [CSI 6 n Report Cursor
Position](https://invisible-island.net/xterm/ctlseqs/ctlseqs.html)
and use it to compute the left prompt's terminal-coordinates (which
are (0, 0) in fish-coordinates).
Unfortunately this doesn't yet work correctly while the terminal
is scrolled. This is probably because the cursor position is wrong
if off-screen. To fix that we could probably record the cursor
position while not scrolled, but it doesn't seem terribly important
(the existing implementation also doesn't get it right).
We still turn off mouse reporting. If we turned it on, it
would be harder to select text in the terminal itself (not fish).
This would typically mean that mouse-drag will alter fish's
selection and shift+mouse-drag or alt+mouse-drag can be used.
To improve this, we could try to synchronize the selection: if parts
of the fish commandline are selected in the terminal's selection,
copy that to fish's selection and vice versa.
Or maybe there is an intuitive criteria, like: whenever we receive a
mouse event outside fish, turn off mouse reporting, and turn it back on
whenver we receive new keyboard input. One problem is that we lose
one event (though we could send it back to the terminal). Another
problem is we would turn it back on too late in some scenarios.
Closes#10932
Commit 01dbfb0a3f (replace writestr() with fwprintf() in reader.cpp,
2016-12-20) accidentally replaced a retry-on-EINTR write with a
non-retrying version. Commit 7f31acbf9b (Prevent fish_title output
from triggering a bel, 2022-02-02) fixed this for some cases but
not all, fix that.
This has been removed, see kitty commit cd92d50a0 (Keyboard protocol:
Remove CSI R from the allowed encodings of the F3 key as it conflicts
with the *Cursor Position Report* escape code, 2022-12-24).
cursor_selection_mode=inclusive means the commandline position is
bounded by the last character. Fix a loop that fails to account
for this.
Fixes d51f669647 (Vi mode: avoid placing cursor beyond last character,
2024-02-14).
This change looks very odd because if the commandline is like
echo foo.
it makes us try to uppercase the trailing period even though that's
not part of word range. Hopefully this is harmless.
Note that there seem to be more issues remaining, for example Vi-mode
paste leaves the cursor in an out-of-bounds odd position.
Fixes#10952Closes#10953
Reported-by: Lzu Tao <taolzu@gmail.com>
If base directories (e.g. $HOME/.config/fish) need to be created,
create them with mode 0700 (i.e. restricted to the owner).
This both keeps the behavior of old fish versions (e.g. 3.7.1) and is
compliant with the XDG Base Directory Specification.
See: https://specifications.freedesktop.org/basedir-spec/0.8/#referencing
[Do NOT cherry-pick to 4.0 - this needs more time to be tested]
fish sometimes needs to capture the output of a command or block of
commands. Examples include fish_prompt or any command substitution
("cmdsubs"). It does this the obvious way: by creating a pipe, using dup2
to replace stdout of the command with the write end of the pipe, and then
reading from the read end into a buffer, until EOF or the command
substitution completes. Importantly, this task also overlaps with waiting
for the process to exit; that is when executing:
set var (some_cmd)
fish needs to both wait on `some_cmd` and ALSO read its output into memory.
This is awkward to do in a portable way in a single thread (though maybe
doable on Linux with pidfd). So we wait and read on different threads.
To make things worse, command substitutions may themselves create
additional command substitutions (recursion, etc). Creating a read thread
for every command substitution would result in excessive threads. So rather
than a thread per cmdsub, we have a single dedicated thread that handles
ALL command substitutions, by multiplexing multiple file descriptors via
select/poll. This is the "fd monitor." You hand it a file descriptor and it
lets you know when it's readable, and then you can read from it (via a
callback). Also, it has a "wakeup" fd: if you write to that then the fd
monitor wakes up, figures out what it has to do, and resumes.
When the command substitution ends, we need to remove the fd from the fd
monitor, because we intend to close it. You might object "the commands in
the cmdsub have all completed so the write end of the pipe has been closed
so the fd monitor can just notice that the pipe is closed" but it's not so:
consider the horrible case of `set var (yes &)` and abandon all hope.
The current mechanism for removing the fd from the monitor is called a
"poke." We tell the fd monitor (through a "control" self-pipe) to
explicitly wake up the item. It then invokes the callback ("pokes") the
item on the dedicated fd monitor thread. The item notices that the command
substitution is complete, and it returns a value meaning "remove me" and
the fd monitor does so. The client thread is stuck waiting for this process
to complete.
So basically removing a fd from the monitor requires a round trip to its
dedicated thread. This is slow and also complicated (Rust doesn't have
futures)!
So let's not do that.
The big idea is to remove this round-trip synchronization. That is, when we
intend to remove the fd from the fd monitor, we _just do it_ and then close
the fd. Use a lock rather than a round-trip to the thread. Crucially that
lock is unlocked while the monitor thread waits in select/poll.
This invites all sorts of races:
1. fish might remove and close the fd right before the monitor polls it. It
will thus attempt to poll a closed fd.
2. fish might remove and close the fd, and then something else opens a file
and receives the same fd. Now the fd monitor will poll an fd that was
never added.
3. fish might remove and close the fd _while the fd monitor is polling it_.
What happens then? (Turns out on macOS we get EBADF, and on Linux the fd is
marked readable).
The Big Idea is that *all of these races are benign*. As long as
poll/select doesn't crash or hang, we don't care *what* it returns, because
the source of truth are the set of items stored in the fd monitor and these
item IDs are never recycled. (This also assumes that it's OK to select/poll
on random file descriptors; there ought to be no side effects).
Not only is this a large simplification since we no longer need that round
trip, it's a substantial performance improvement as well. The
"aliases.fish" benchmark goes from 164 to 154 msec on my Mac, and from 124
to 112 msec on my Linux machine - nearly 10%.
Add some tests to verify our assumptions about the behavior of closing or
replacing a file descriptor during poll. But even if these fail, all we
care about is that poll/select doesn't crash or hang.
FdMonitor is used to monitor a set of file descriptors and invoke a callback
when one becomes readable. Prior to this commit, they coudl also have the
callback invoked on timeout. fish used to use this feature but no longer does;
remove it.
It is, as the name implies, unused - it became SIGSYS, which we
already check.
Since it is entirely undefined on some architectures it causes a build
failure there, see discussion in #10633
The libc crate has a bug on BSD where WEXITSTATUS is not an 8-bit
value, causing assertion failures.
Any libc higher than our 0.2.155 would increase our MSRV, see libc
commit 5ddbdc29f (Bump MSRV to 1.71, 2024-01-07), so we want to
woraround this anyway. It's probably not worth using a patched
version of libc since it's just one line.
While at it, tighten some types I guess.
Upstream fix: https://github.com/rust-lang/libc/pull/4213Closes#10919
__fish_cancel_commandline was unused (even before) and has some issues
on multiline commandlines. Make it use the previously active logic.
Closes#10935
* Pass path to install()
It was dirty that it would re-get $HOME there anyway.
* Import wcs2osstring
* Allow installable builds to use a relocatable tree
If you give a path to `--install`, it will install fish into a
relocatable tree there, so
PATH/share/fish contains the datafiles
PATH/bin/fish contains the fish executable
PATH/etc/fish is sysconf
I am absolutely not sold on that last one - the way I always used
sysconfdir is that it is always /etc. This would be easy to fix but
should probably also be fixed for "regular" relocatable builds (no
idea who uses them).
An attempt at #10916
* Move install path into "install/" subdir
* Disable --install harder if not installable
Ever since 149594f974 (Initial revision, 2005-09-20), we move the
cursor to the end of the commandline just before executing it.
This is so we can move the cursor to the line below the command line,
so moving the cursor is relevant if one presses enter on say, the
first line of a multi-line commandline.
As mentioned in #10838 and others, it can be useful to restore the
cursor position when recalling commandline from history. Make undo
restore the position where enter was pressed, instead of implicitly
moving the cursor to the end. This allows to quickly correct small
mistakes in large commandlines that failed recently.
This requires a new way of moving the cursor below the command line.
Test changes include unrelated cleanup of history.py.
Commit 8bf8b10f68 (Extended & human-friendly keys, 2024-03-30) stopped
ctrl-c from exiting without a motivation. Unfortunately this was
only noticeable on terminals that speak the kitty keyboard protocol,
which is probably no one had noticed so far.
Closes#10928
This is allowed
time a=b echo 123
but -- due to an oversight in 3de95038b0 (Make "time" a job prefix,
2019-12-21) -- this is not allowed:
not time a=b echo 123
Instead, this one one works:
not a=b time echo 123
which is weird because without the "not" this would run "/bin/time".
It seems wrong that "not" is not like the others. Swap the order
for consistency.
Note that unlike "not", "time" currently needs to come before variable
assignments, so "a=b time true" is disallowed. This matches zsh. POSIX
shells call "/bin/time" here. Since it's ambiguous, erroring out seems
fine. It's weird that we're inconsistent with not here but I guess
"command not" is not expected to have subtly different behavior.
Closes#10890
This was added accidentally in 971d257e67 (Port AST to Rust,
2023-04-02). It does not seem to be causing an observable effect
(although I didn't try hard).
The values we would try are:
xterm-256color, xterm, ansi, dumb
This is a pretty useless list, because systems without
"xterm-256color" but with "ansi" basically don't exist,
and it is very likely that the actual terminal is more
xterm-compatible than it is ansi.
So instead we just use our xterm-256color definition, which has a high
likelihood of being basically correct.
This is fairly subtle.
When installable, and we either can't find the version file or it is
outdated, we ask the user to confirm installation (just like `--install`).
We do that only if we are really truly interactive (with a tty!) to
avoid `fish -c` running into problems.
This check could be tightened even more, because currently:
```fish
fish -ic 'echo foo'
```
asks, while
```fish
fish -ic 'echo foo' < /dev/null
```
does not.
`fish -c` will still error out if it can't find the config, but it
will just run if it is out of date.
fish will print messages for some jobs when they exit abnormally, such as
with SIGABRT. If a job exits abnormally inside the prompt, then (prior to
this commit) fish would print the message and re-trigger the prompt, which
could result in an infinite loop. This has existed for a very long time.
Fix it by reaping jobs after running the prompt, and NOT triggering a
redraw based on that reaping. We still print the message but the prompt is
not executed.
Add a test.
Fixes#9796
This built on my test system, might be version differences.
(it's also not enough to make it *work*, but a necessary step)
This reverts commit 6fded249cd.
Commit 29dc30711 (Insert some completions with quotes instead of
backslashes, 2024-04-13) wrongly copmletes
$ cat ~/space
to
$ cat '~/path with spaces'
Today completions can be either replacing or appending. We never quote
(but backslash-escape) appending completions (unless they "append"
to an empty token). We always quote replacing completions. The
assumption in this part of the code is that replacing completions
can be quoted without changing meaning.
This assumption is wrong for tildes. For the backslash-escaping code
path, we take care of this edge case via a special DONT_ESCAPE_TILDES
flag. However that flag does not take effect when using quotes for
escaping. Fix that.
Unfortunately, e97a4fab7 (Escape : and = in file completions,
2024-04-19) introduced a (hopefully temporary) code clone in
escape_separators, which made added an extra step to debugging here.
When built with the default "installable" feature, the data files (share/) are
included in the fish binary itself.
Run `fish --install` or `fish --install=noconfirm` (for
non-interactive use) to install fish's data files into ~/.local/share/fish/install
To figure out if the data files are out of date, we write the current version
to a file on install, and read it on start.
CMake disables the default features so nothing changes for that, but this allows installing via `cargo install`,
and even making a static binary that you can then just upload and have extract itself.
We set $__fish_help_dir to empty for installable builds, because we do not have
a way to generate html docs (because we need fish_indent for highlighting).
The man pages are found via $__fish_data_dir/man
This now allows:
- Same argument (`random 5 5`)
- Swapped ends (`random 10 2`)
- One possibility (`random 0 5 4`)
This makes it easier to use with numbers generated elsewhere instead
of hard-coded, so you don't need to check as much before running it.
Fixes#10879
We don't really care if the process has a custom handler installed, we
can just set it to default.
The one we check is SIGHUP, which may be given to us via `nohup`.
This saves ~30 syscalls *per process* we spawn, so:
```fish
for f in (seq 1000)
command true
end
```
has ~30000 fewer rt_sigaction calls. These take up about ~30% of the
total time spent in syscalls according to strace.
We could also compute this set once at startup and then reuse it.
exists_no_autoload() wrongly thinks that tombstoned functions can be
autoloaded; fix that.
While at-it replace the use of get_props() with something simpler.
Co-authored-by: Himadri Bhattacharjee
Closes#10873
The [disambiguate flag](https://sw.kovidgoyal.net/kitty/keyboard-protocol/#disambiguate) means that:
> In particular, ctrl+c will no longer generate the SIGINT signal,
> but instead be delivered as a CSI u escape code.
so cancellation only works while we turn off disambiguation.
Today we turn it off while running external commands that want to
claim the TTY. Also we do it (only as a workaround for this issue)
while expanding wildcards or while running builtin wait.
However there are other cases where we don't have a workaround,
like in trivial infinite loops or when opening a fifo.
Before we run "while true; end", we put the terminal back in ICANON
mode. This means it's line-buffered, so we won't be able to detect
if the user pressed ctrl-c.
Commit 8164855b7 (Disable terminal protocols throughout evaluation,
2024-04-02) had the right solution: simply disable terminal protocols
whenever we do computations that might take a long time.
eval_node() covers most of that; there are a few others.
As pointed out in #10494, the logic was fairly unsophisticated then:
it toggled terminal protocols many times. The fix in 29f2da8d1
(Toggle terminal protocols lazily, 2024-05-16) went to the extreme
other end of only toggling protocols when absolutely necessary.
Back out part of that commit by toggling in eval_node() again,
fixing cancellation. Fortunately, we can keep most of the benefits
of the lazy approach from 29f2da8d1: we toggle only 2 times instead
of 8 times for an empty prompt.
There are only two places left where we call signal_check_cancel()
without necessarily disabling the disambiguate flag
1. open_cloexec() we assume that the files we open outside eval_node()
are never blocking fifos.
2. fire_delayed(). Judging by commit history, this check is not
relevant for interactive sessions; we'll soon end up calling
eval_node() anyway.
In future, we can leave bracketed paste, modifyOtherKeys and
application keypad mode turned on again, until we actually run an
external command. We really only want to turn off the disambiguate
flag.
Since this is approach is overly complex, I plan to go with either
of these two alternatives in future:
- extend the kitty keyboard protocol to optionally support VINTR,
VSTOP and friends. Then we can drop most of these changes.
- poll stdin for ctrl-c. This promises a great simplification,
because it implies that terminal ownership (term_steal/term_donate)
will be perfectly synced with us enabling kitty keyboard protocol.
This is because polling requires us to turn off ICANON.
I started working on this change; I'm convinced it must work,
but it's not finished yet. Note that this will also want to
add stdin polling to builtin wait.
Closes#10864
Moving the "make empty ToLowercase iterator" logic to within the
`unwrap_or_else()` instead of always generating it brings most of the speedup;
unrolling the recursive call brings in the rest.
Using `c.is_uppercase()` instead of getting the iterator and checking if the
first (and only) lowercase letter of the sequence is the same as the original
input is 5-8x faster (measured via criterion against `/usr/share/dict/words`).
(Additional benefit of forcibly inlining the now iterator-based comparison not
taken into account; this necessitated changing from a closure to a local
function as the inline attribute on closures is not yet supported with the
stable compiler toolchain.)
This is still suboptimal because we are allocating a vector of indices to be
removed (but allocation-free in the normal case of no duplicates) but
significantly better than the previous version of the code that duplicated the
strings (which are larger and spread out all over the heap).
The ideal code (similar to what we had in the C++ version, iirc) would look like
this, but it's not allowed because the borrow checker hates you:
```
fn unique_in_place_illegal(comps: &mut Vec<Completion>) {
let mut seen = HashSet::with_capacity(comps.len());
let mut idx = 0;
while idx < comps.len() {
if !seen.insert(&comps[idx].completion) {
comps.remove(idx);
continue;
}
idx += 1;
}
}
```
If a semicolon-delimited list of CSI parameters contained an (invalid) long
sequence of ascii numeric characters, the original code would keep multiplying
by ten and adding the most recent ones field until the `params[count][subcount]`
u32 value overflowed.
This was found via automated fuzz testing of the `try_readch()` routine against
a corpus of some proper/valid CSI escapes.
This lets us call into the entirety of the prior `readch()` with an exhaustible
input stream without panicking on the `unreachable!()` call. The previous
functionality is kept under the old name by calling `try_readch()` with the
`blocking` parameter set to `true` (100% same behavior as before).
While the `try_readch(false)` entrypoint isn't used directly by the current fish
codebase, it is required in order to automate input reader tests without the
overhead and complexity of running the test harness in a tty emulator emulator
like pexpect or tmux, which moreover necessitates out-of-process testing – which
is incompatible with most perf-guided testing harnesses.
I hope to be able to upstream harness integrations using this entry point in the
near future.
`print_help` is a hacky-wacky function used to support the `--help` command
of `fish_key_reader` and others. The Rust version panics on an error; fix
that and make it print more useful help messages.
I'm guessing this was missed in the port because there were comments referencing
using a hash set to perform the deduplication but there was no hashset. (The
TODO was added later.)
This caught an incorrect description for process/job exit handlers for ANY_PID
(now removed) which has been replaced with a message stating the handler is for
any process exit event.
The previous approach of "treat this field as an `Option<NonZeroU32>` and
remember to check `p.has_pid()` before accessing it" was a mix of C++ and rust
conventions and led to some bugs or incorrect behaviors.
* `jobs -p` would previously print both the (correct) external pid and the
(incorrect) internal value of `0` if a backgrounded command contained a
fish function (e.g. `function foo; end; cat | foo &; jobs`)
* Updating/calculating job cpu time and usage was incorrectly including all of
fish's cpu usage/time for each function/builtin member of the job pipeline.
Closes#10832
ctrl-r ctrl-s ctrl-s
Attemps to go before the beginning and asserts out. Instead refuse to
do that.
(there's some weirdness where it can reduce the pager to the first
entry if you keep pressing, which I haven't found yet, but that's better than *crashing*)
These are another way to spell the same thing that doesn't match what
`bind` would print.
They're also not documented and tested thoroughly.
Since they are just small shortcuts and unreleased we can just remove
them.
Fixes#10845
Given "printf %18446744073709551616s", we parse the number only in
the printf crate, which tells us that we overflowed somwhere (but
not where exactly).
We were previously printing the internal `INVALID_PID` value (since removed),
which was a meaningless `-2` constant, when there was no pgid associated with a
job.
This PR changes that to `-` to indicate no pgid available, which I prefer over
something like `0` or `-1`, but will cause problems for code that is hardcoded
to convert this field to an integral value.
