Spent a bit in a deep dive into how to handle this and honestly the
situation is rather unfortunate. The core problem is that when we have a
panic anywhere we need to tear down the app, and we'd like to do that as
cleanly as possible, avoiding throwing any other panics along the way if
possible.
We've been seeing a number of panics being reported which are
nonsensical, seemingly pointing to being a fallout panic from a worker
thread panic-ing, at which point we would write multiple panics to the
panic file, and we could possibly upload either both or the wrong panic
causing a wild goose chase. Unfortunately I've been entirely unable to
reproduce the specific panic we've been seeing but I was able to read
through the code responsible and confirm that under specific situations
a panic on one worker can cause another worker or the main thread to
also panic.
An easy solution to this is just to ignore any panics after the first
one. I'm thinking that *hopefully* we can trust the first panic to reach
the panic hook first so that the flag doesn't accidentally filter out
the panic we actually care about.
That being said we were expecting that to have already been the case
about which panic gets written to the panic file first, the first one in
the file being the one we upload, which doesn't seem to have been the
case. I'm hoping it was IO silliness causing that and that the flag
shouldn't be race-y, however this is still a shot in the dark. 🤞
As for cleanly shutting down, there's not really much we can do. One
thread physically cannot cause another to unwind without somehow sending
a message which isn't super useful. The only way for a thread to shut
down all threads and the process is to go nuclear and abort/exit the
process. This will never unwind other threads, effectively having the
same effect on those threads as compiling with `panic = "abort"` would.
With some (mis)use of `std::panic::resume_unwind` we can at least say
that for whatever thread actually panic-ed we will unwind, and any other
threads that panic as a result will probably get at least partway
through unwinding. This is weird, almost a combination of panic
rewinding and aborting, and may actually be worse than just biting the
bullet and aborting immediately.
I'm really not a fan of where I've ended up but it does seem to at the
very least an improvement. The main question in my mind at this point is
whether it would be better to attempt to unwind what we can or go all in
on abort. I'd love some input on that.
Release Notes:
- Improved panic reporting when a background thread panics.
Linear:
https://linear.app/zed-industries/issue/Z-2578/zed-launches-very-slow-for-user
I was searching for the cause of a slow startup time reported in the
above issue, and I don't think I found it, but I did find two very
noticeable slow code paths while profiling, and fixed them.
### Notes
1. When starting the JSON language server, we provide it with a JSON
schema for our settings. For the `theme` setting, the JSON schema needs
to read all of the themes in the registry, to generate a list of valid
theme names. Previously, as part of this, we were deserializing each
theme from JSON, which took a lot of CPU. Now, we don't do that.
2. When an FS event occurs within a git repository, we reload the git
status for all entries in that git repository. Previously, we did that
via a separate `libgit2` call per FS entry (including ignored entries,
so many thousands in the case of the `zed` repo). Now we do one
`libgit2` call, asking for all of the statuses. Git carries an index of
all of the files with statuses, so this is fast.
Release Notes:
- Improved the the performance of starting up a JSON language server.
- Improved the performance of handling changes to git repositories, such
as changing branches or committing.
Closes https://github.com/zed-industries/community/issues/75
Closes https://github.com/zed-industries/community/issues/1749
The PR
* changes keybindings for `Editor && mode == auto_height` context:
before, `alt-enter` and `alt-shift-enter` added new lines in such
editors, including the one from buffer search.
New bindings are the same as in `Editor && mode == full` context.
* adds `search::SelectAllMatches` action and binds it to `Alt + Enter`
by default, to select all matches of a buffer search
The behavior mimics VSCode: we do not move the screen even if all
selections are out of the visible range (Cmd+G will navigate there) and
allow reselecting the results from both pane and search field, as long
as the search is not dismissed.
Release Notes:
- Added `search::SelectAllMatches` (`Alt + Enter` default) action to
place carets and select all buffer search results
([#75](https://github.com/zed-industries/community/issues/75),
[#1749](https://github.com/zed-industries/community/issues/1749)).
We're finally doing the thing.
TODO:
- [x] Choose an approach
- Decided to add a new element just for the pane axis, containing a
slimmed down copy of the flex code.
- [x] Wire through callbacks and pointers so that data goes where it
needs to
- [x] Do the flex juggling math on resize
- [x] Update the flexes when updating the split tree
- [x] Restore the active_pane_magnification setting
- [x] Serialize an axis' flexes
Release Notes:
- Made the center pane group splits resizable. Note that resizing is
disabled if the `active_pane_magnification` setting is changed from
default.
