[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.
Instead of hardcoded 230px margin.
This also makes the ToC only take up a third of the screen when
narrow, and lets you scroll the rest.
Without, you'd have to scroll past the *entire* ToC, which is awkward
Remaining issue is the search box up top. Since this disables the one
in the sidebar once the window gets too narrow, that one is important,
and it isn't *great*
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
I forgot that 610338cc70 (On undo after execute, restore the cursor
position, 2024-12-21) would cause a fallout to tests:
It makes us reuse in another place our usual cursor-movement sequences.
This causes failures like this (linebreaks added for readability):
Testing file pexpects/bind.py:Failed to match pattern: (?:\r\n|\x1b\[2 q)[^\n]*def abc\r\n
bind.py:45: timeout from expect_prompt(TO_END + "def abc\r\n") # emacs transpose words, default timeout: no delay
Escaped buffer:
\x1b[?2004h\x1b[>4;1m\x1b[=5u\x1b=\rprompt 2>echo \rprompt 2>echo abc \rprompt 2>echo def abc\r
prompt 2>echo def abc\x1b[?2004l\x1b[>4;0m\x1b[=0u\x1b>\x1b]133;C\x07def abc\r\n\x1b]133;D;0\x07\x1b[?25h⏎
\r⏎ \r\rprompt 3>\x1b[?2004h\x1b[>4;1m\x1b[=5u\x1b=
It seems that we don't print anything where we should print something
like "\r\n" or "\e[2 q" to move the cursor below the command line.
I haven't gotten to the bottom of this but it might be related to
terminfo. Once we get rid of that, we can unconditionally print
our canonical movement sequences.
This issue seems to only affect tests, since fish operates fine in
a sourcehut CI system. Let's ignore it for now.
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 reverts commit 27c7578760.
dust generates its own completions (which are shipped in the wrong spot
in the Debian packages, but which are also more up-to-date).
Closes#10922.
The apple-codesign crate has a fairly aggressive MSRV policy, and the
compiler itself still targets 10.12 which is well below the minimum
version of macOS for aarch64. Just use stable.
