Instead of a single 'job read' callback, job control consumers need to provide
callbacks for "stdout read", "stderr read" and "exit". For vimscript, the
JobActivity autocommand is still used to handle every job event, for example:
```vim
:let srv1_id = jobstart('netcat-server-1', 'nc', ['-l', '9991'])
:let srv2_id = jobstart('netcat-server-2', 'nc', ['-l', '9991'])
function JobEvent()
" v:job_data[0] = the job id
" v:job_data[1] = the event type, one of "stdout", "stderr" or "exit"
" v:job_data[2] = data read from stdout or stderr
if v:job_data[1] == 'stdout'
let str = 'Message from job '.v:job_data[0].': '.v:job_data[2]
elseif v:job_data[1] == 'stderr'
let str = 'Error message from job '.v:job_data[0].': '.v:job_data[2]
else
" Exit
let str = 'Job '.v:job_data[0].' exited'
endif
call append(line('$'), str)
endfunction
au JobActivity netcat-server-* call JobEvent()
```
And to see messages from 'job 1', run in another terminal:
```sh
bash -c "while true; do echo 123; sleep 1; done" | nc localhost 9991
```
RStream will be the main way Neovim receives asynchronous messages, so it is
best to have a specialized EventType for it. A new flag parameter was added to
`rstream_new` which tells the RStream instance to defer event handling for later
with KE_EVENT instead of handling it directly from libuv callback.
As of v0.11.23 libuv's uv_timer_cb, uv_async_cb, uv_prepare_cb, uv_check_cb and
uv_idle_cb no longer require a status parameter (this went unused in the first
place).
Bump third-party dependency `libuv` up to 0.11.23 and remove the extra
parameters from the callbacks.
K_EVENT/KE_EVENT are used to signal any loop that reads user input(scattered
across normal.c edit.c , ex_getln.c and message.c) of asynchronous events that
were not initiated by the user.
Representing non-user asynchronous events as special keys has the following
advantages:
- We reuse the normal vim redrawing code. As far as the rest of the code in
edit.c/normal.c is concerned, it's just the user pressing another key.
- Assume less about vim tolerance for "out-of-band" modifications to its
internal state.
- We still have a very complex codebase and it's hard to predict what bugs may
be introduced by these changes. With this we implement asynchronicity in a way
that will be more "natural" to the editor and has less chance of causing
unpredictable behavior.
As the code is refactored, we will be able to treat user input as an 'event
type' and not the other way around(With this we are treating arbitrary events as
a special case of user input).
- Add a job control module for spawning and controlling co-processes
- Add three vimscript functions for interfacing with the module
- Use dedicated header files for typedefs/structs in event/job modules
- Add a new macro to klist.h: kl_empty()
The whole point of abstract data structures is to avoid reimplementing
common actions. The emptiness test seems to be such an action.
- Add a new function attribute to func_attr.h: FUNC_ATTR_UNUSED
Some of the many functions created by the macros in klist.h may end up not
being used. Unused functions cause compilation errors as we compile with
-Werror. To mark those functions as possibly unused we can use the
FUNC_ATTR_UNUSED now.
- Pass `Event` by value
`Event` is such a small struct that I don't think we should allocate heap space
and pass it by reference. Let's use the stack and memory cache in our favor
passing it by value.
This removes all signal handling code from os_unix.c to os/signal.c. Now signal
handling is done like this:
- Watchers for signals are registered with libuv default event loop
- `event_poll` continuously calls `poll_uv_loop` to produce events until it
receives user input, SIGINT or a timeout
- Any signals received in `poll_uv_loop` will push events to a queue that is
drained and processed by `event_poll`
Signals aren't handled directly in the libuv callback to avoid recursion in the
event loop(which isn't supported by libuv).
The same principle will apply to other events in the future: Push to a queue
from a libuv callback and drain it from `event_poll`
This feature was accidentally removed when doing the initial import from vim. It
makes vim use pipes instead of temporary files for filtering buffers through
shell commands.
I found that this was missing when looking for references of
SHELL_READ/SHELL_WRITE outside mch_call_shell`.
When `mch_call_shell` is reimplemented on top of libuv process management
facilities, pipes will always be used for communication with child processes so
it makes sense to enable the feature permanently.
The functions `mch_inchar`, `mch_breakcheck`, `mch_char_avail` were
reimplemented on top of libuv. Here's how it works:
- When Neovim needs to wait for characters, it will transfer control to libuv
event loop.
- When the libuv event loop gets user input, it will transfer control back to
Neovim
- Neovim uses the `input_read` function to get the actual data read by libuv.
With this scheme its possible to keep Neovim single-threaded while enjoying the
benefits provided by libuv.
This commit leaves SIGWINCH broken for now
