I was running `hg log --config pager.pager=sp` and pressing space without
releasing it. After about 10k lines sp appears to deadlock. It seems sp uses a
single thread for both reading terminal events and sending wake events and it
sends too many wake events without reading the events.
Failing to write to the wake pipe due to EWOULDBLOCK does not seem to be a
problem - there are enough events in the pipe to wake up the other side.
Therefore let's just make wake_pipe_write nonblocking and treat EWOULDBLOCK as
a success.
Context: The stuck thread looks like:
#0 0x00007f32671ee237 in write () from /usr/lib/libc.so.6
#1 0x000055c466022823 in std::sys::unix::fd::FileDesc::write () at src/libstd/sys/unix/fd.rs:96
#2 std::sys::unix::net::Socket::write () at src/libstd/sys/unix/net.rs:276
#3 <&std::sys::unix::ext::net::UnixStream as std::io::Write>::write ()
at src/libstd/sys/unix/ext/net.rs:597
#4 <std::sys::unix::ext::net::UnixStream as std::io::Write>::write ()
at src/libstd/sys/unix/ext/net.rs:582
#5 0x000055c465d010a6 in termwiz::terminal::unix::UnixTerminalWaker::wake (self=0x7ffe6bd32de0)
at /home/quark/.cargo/git/checkouts/wezterm-6425bab852909cc8/ef1b836/termwiz/src/terminal/unix.rs:278
#6 0x000055c465a6c64b in streampager::event::EventSender::send_unique (self=0x7ffe6bd32dd0, event=...,
unique=0x7ffe6bd32de8) at src/event.rs:66
#7 0x000055c465a7e65a in streampager::display::start (term=..., term_caps=..., events=..., files=...,
error_files=..., progress=..., mode=streampager::config::FullScreenMode::Auto) at src/display.rs:295
#8 0x000055c465a453a7 in streampager::Pager::run (self=...) at src/lib.rs:171
#9 0x000055c465a0aced in sp::open_files (args=...) at src/bin/sp/main.rs:170
#10 0x000055c465a08e4f in sp::main () at src/bin/sp/main.rs:25
This commit adds some plumbing for describing the cursor shape
(block, line, blinking etc) and visibility, and feeds that through
the mux and render layers.
The renderer now knows to omit the cursor when it is not visible.
This isn't complete but begins the process of extracting
the embedding application configuration into a trait provided
by the application rather than passing the values in at
construction.
This allows the application to change configuration at
runtime.
The first option to handle this is the scrollback size.
The defaults are pretty neutral. You can get a little more fancy
with something like this:
```
[colors.tab_bar]
background = "#0b0022"
[colors.tab_bar.active_tab]
bg_color = "#2b2042"
fg_color = "#c0c0c0"
[colors.tab_bar.inactive_tab]
bg_color = "#1b1032"
fg_color = "#808080"
[colors.tab_bar.inactive_tab_hover]
bg_color = "#3b3052"
fg_color = "#909090"
italic = true
```
This is a bit of a large commit because it needed some plumbing:
* Change mux creation to allow deferring associating any domains,
and to change the default domain later in the lifetime of the
program
* De-bounce the empty mux detection to allow for transient windows
during early startup
* Implement a bridge between the termwiz client Surface and the
frontend gui renderer so that we can render from termwiz to
the gui.
* Adjust the line editor logic so that the highlight_line method
can change the length of the output. This enables replacing
the input text with placeholders so that we can obscure password
input
I noticed while scrolling `emoji-test.txt` that some of the combined
emoji sequences rendered very poorly. This was due to the unicode
width being reported as up to 4 in some cases.
Digging into it, I discovered that the unicode width crate uses a
standard calculation that doesn't take emoji combination sequences
into account (see https://github.com/unicode-rs/unicode-width/issues/4).
This commit takes a dep on the xi-unicode crate as a lightweight way
to gain access to emoji tables and test whether a given grapheme is
part of a combining sequence of emoji.
I've noticed this off and on for a while, and thought it was something
fishy with my shell dotfiles.
Tracing through I found that the final byte in the "Face with head
bandage" emoji 🤕 U+1F915 was being interpreted as the MW control
code and causing the vt parser to jump out of the OSC state.
The solution for this is to hook up proper UTF-8 processing in the
same way that it is applied in the ground state.
Since we don't have enough bits to introduce new state values (we're
pretty tightly packed in the 16 bits available), I've introduced a
memory of the state to which the utf8 parser needs to return once
a complete sequence is detected.
I thought that I'd broken something with the DEL processing in vim with
the new frontend but it turned out that the other frontend was emitting
BS always and that I'd actuall unbroken passing DEL through and that
other layers were translating DEL into an application cursor mode output
for DEL that emits a totally different sequence.
This diff preserves DEL and disables that other sequence.
Will follow up with some explicit configuration to control this
behavior, but in the short term, the default behavior should be much
closer to what people actually want and expect!
refs: https://github.com/wez/wezterm/issues/52
This is still a bit rough because the terminal parser doesn't
understand the pixel sizes, so it relies on the hard coded
cell dimensions being accurate.
This enables using large OSC buffers in a form that we can publish
to crates.io without blocking on an external crate. Large OSC
buffers are important both for some tunnelling use cases and for
eg: iTerm2 image protocol handling.
Add a convenience function to the escape parser that, like `parse_first`,
matches only the first escape sequence, but instead collects all matching
actions.
It's taking a while for https://github.com/jwilm/vte/pull/20 to get
merged, so point to my branch directly while I build out some
tunneled mux protocol escape sequences.
I'll need to fork vte on crates.io if vte doesn't merge the PR
before the next termwiz crate bump.
The `flush_pending_attr` method does lots of unnecessary comparisons.
If the attributes have changed, then it works by resetting the
attributes and then setting new values. There's no need to emit the
codes for exiting modes.
It also doesn't support double underscore or rapid blink in the cases
where the terminfo capabilities are used, as these capabilities can't
express these attributes. Fall back to CSI sequences when these
attributes are requested.
Changing the terminal attributes (bold, underline, etc.) involves
emitting the `exit_attribute_mode` or SGR reset sequence. This also
resets the colors back to their defaults. If this happens when the
foreground or background colors haven't changed, set the colors again.
These codes are used to change the color palette, but if the `?`
string is used in place of a color spec, then we must respond with
the current color value string for that palette entry, so lets
implement that!
We need to set the terminal to blocking mode when we want poll_input
to block forever, otherwise we immediately receive WouldBlock error
response and quit the program.
This one has been bugging me for a while; we now know when we've
wrapped a line and can join it without a line break when double-clicking
to select a word.
This commit introduces a wrapped attribute to help record this
information, which could potentially help with when it comes
to looking at nicer resize behavior in refs: https://github.com/wez/wezterm/issues/14
When repainting the screen, we must be sure to set the cursor
to the right shape.
While the repaint is happening, hide the cursor to prevent
the cursor jumping around while the repaint happens.
Extract the code that builds a `ProbeHintsBuilder` from the environment to a
separate `ProbeHintsBuilder` constructor. This allows callers to re-use the
environment-based `ProbeHintsBuilder`, but override other aspects of
`ProbeHints`, e.g. to disable mouse handling.
When repainting a surface, we optimize for the case where lines are simple text
by combining the the `Change::Text` for the end of the previous line and the
start of the next line into a single `Change`.
The assumption about relying on automatic margins is incorrect. We can't rely
on them, as they might be disabled, and in any case they are no use if the
previous line was shorter than the full width of the screen.
This results in the lines appearing joined together on a single line. This is
evidenced in the existing tests where `"hel\nw"` becomes `"helw"` on a full
repaint.
The solution is to always inject a real CRLF by adding a `CursorPosition` change.
This replaces any CRLF that may have been swallowed by the `Surface` when it
added the original changes.
`input_parser.decode_input_records` might not add anything to the input queue,
e.g. if the input event is one that is being ignored. In this case, we must
loop round and wait again for more input.
Remove the dance for appeasing the borrow checker. The borrow checker can
be appeased by borrowing the `input_queue` field directly.
Application cursor keys are a separate set of encodings that applications can
ask the terminal to produce for cursor keys.
Unfortunately, PuTTY generates these for shift-modified cursor keys. If an
application is to distinguish between normal and shift-modified cursor keys
on PuTTY then it will need to be able to distinguish between normal and
application cursor keys.
Add new `KeyCode` variants for application cursor keys.
Add two new `Change` variants: `ScrollRegionUp` and `ScrollRegionDown`, which
scroll part of the screen up or down by a number of lines.
On Unix, these are implemented using the `change_scroll_region` and
`parm_index`/`parm_rindex` terminfo capabilities if available. If `parm_index`
or `parm_rindex` are not available, but `scroll_forward` or `scroll_reverse`
are, then these are used repeatedly to get the same effect.
On Windows, these are implemented using `ScrollConsoleScreenBuffer`.
If the cursor has been made invisible (by setting the shape to `Hidden`),
then in order for it to be visible when it is reset to the default shape,
it must be made visible. This can be done by rendering the `CursorVisible`
sequence.
Where supported, the `CursorNormal` sequence can be rendered to do
both `CursorVisible` and `ResetCursorStyle` in one go.
This was an off-by-one issue when using scroll margins just shorter
than the screen and when scrollback is enabled.
Added a unit test to verify the behavior.
I've had mixed results with esctest; the IRM and cursor save/restore
tests fail for me in terminal.app, iterm2 and xterm, and fail in the
same way on wezterm, so I'm not sure if I'm not running those tests
correctly. However, they did encourage the discovery of some other
real issues in the wezterm emulation.
These were showing up as unrecognized sequences on my mac, and I
wondered what they were. We now parse them and do nothing with
them, other than print out the parsed form :)
I wanted to see how much work remains to enable iterm2 image
display; one of the blockers was a limit in the size of the
buffer in the vte crate, which has been removed in my fork
of vte.
As part of testing our ability to absorb that data, I found
a couple of issues with applying the image cells to the display,
so this commit also takes care of that.
We still don't have code to connect the cell image data to the
opengl render layer.
The saturating_sub was going to be an attempt at the look-back alluded
to in the TODO comment, but I decided not to implement that in this
iteration.
The effect of this was that double clicking the `m` in a sequence like
` master` would only select the `m` instead of the `master` string.
previously this just used unicode word segmentation rules, but that
is insufficient for most technical users.
Change this to look for sequences that are non-whitespace and not
enclosed by bracket/quote delimiters. This allows selecting file names
with a double click, which was my main issue.
We don't yet have any code to render them, and the vte parser seems
to truncate incoming image sequences ~1kb in size, so more work is
needed to make this useful.
