Also:
- Remove NO_CONSOLE_INPUT/NO_CONSULE preprocessor conditionals
- Remove ctrl_c_interrupts variable, check for mapped_ctrl_c directly in
process_interrupts()
- Move ui_inchar profiling to input_poll which is where Nvim blocks for input.
Two new functions, `event_enable_deferred()`/`event_disable_deferred()` have to
be called by code that is capable of handling asynchronicity. User-dialog states
like "press ENTER to continue" or the swap file confirmation no longer will
generate K_EVENT.
Commit @45525853d352 removed usage of the `job_write_cb` for closing stdin due
to a memory error, but that doesn't work anymore because `job_close_in` closes
stdin immediately, possibly trimming input data before it is fully written.
Since most memory issues with jobs have been fixed, re-add the `job_write_cb`
call to ensure stdin is only closed when it should. Also add tests for scenarios
where using the callback makes a difference.
- Add input_teardown/signal_teardown to take care of closing signal/stdin
handles.
- Call those functions in event_teardown, and ensure there are no active handles
by entering an infinite loop when there are unclosed handles(think of this as
an assertion that can't go unoticed on travis).
- Move event_teardown call to the end of mch_exit. That is required because
event_poll may still be called in that function.
Commit @264e0d872c("Remove automatic event deferral") removed the immediate
event queue because event deferral now had to be explicit. The problem is that
while some events don't need to be deferred, they still can result in recursive
`event_poll` calls, and recursion is not supported by libuv. Examples of those
are msgpack-rpc requests while a server->client request is pending, or signals
which can call `mch_exit`(and that will result in `uv_run` calls).
To fix the problem, this reintroduces the immediate event queue for events that
can potentially result in event loop recursion. The non-deferred events are
still processed in `event_poll`, but only after `uv_run` returns.
- Rename WriteData to WRequest
- Inline uv_write_t into WRequest, avoiding an extra allocation
- Manage WBuffer/WRequest instances using klib memory pools
Use a timer to periodically compare the current HR time against the HR time of
when `job_stop` was called. After 1 second, send SIGTERM, after 2 seconds, send
SIGKILL. The timer is only active when there's at least one `job_stop` call
pending.
Passing NULL as the callback for stdout/stderr will result in job_start ignoring
stdout/stderr, respectively. A 'writable' boolean argument was also added, and
when false `job_start` will ignore stdin.
Also, refactor os_system to allow passing NULL as the `output` argument.
Commit @709685b4612f4 removed the close_job_* calls when uv_spawn fails because
of memory errors when trying to cleanup unitialized {R,W}Stream instances, but
the uv_pipe_t instances must be closed because they are added to the event loop
queue by previous `uv_pipe_init()` calls
The `vim_feedkeys` must be deferred because it can potentially free the buffer
passed to `os_inchar`(which in turns calls `vim_feedkeys` indirectly).
The new `vim_input` function can be used to emulate user input(Since it does not
mess with the typeahead, it is safe to execute without deferring).
- Extract `process_interrupts` out of `convert_input`
- Instead of waiting for os_breakcheck/os_inchar calls, call `convert_input`
and `process_interrupts` directly from the read callback in input.c.
- Remove the `settmode` calls from `job_wait`. Now that interrupts are
processed in the event loop, there's no need to set the terminal to cooked
which introduces other problems(ref 7.4.427)
It's possible that a child process won't close it's standard streams, even after
it exits. This can be evidenced with the "xclip" program:
:call system('xclip -i -selection clipboard', 'DATA')
Before this commit, the above command wouldn't return, even though the xclip
program had exited. That is because `xclip` wasn't closing it's stdout/stderr
streams, which would block pending_refs from ever reaching 0.
Now the job.c module was refactored to ensure all streams are closed when the
uv_process_t handle is closed.
All files under the os, api and msgpack_rpc directories have -Wconversion
automatically applied. CONV_SOURCES is also checked for missing files(when
renaming, for example)
A pattern that is becoming common across the project is to poll for events until
a certain condition is true, optionally passing a timeout. To address this
scenario, the event_poll_until macro was created and the job/channel/input
modules were refactored to use it on their blocking functions.
This is how asynchronous events are currently handled by Nvim:
- Libuv event loop is entered when Nvim blocks for user input(os_inchar is
called)
- Any event delivered by libuv that is not user input is queued for processing
- The `K_EVENT` special key code is returned by os_inchar
- `K_EVENT` is returned to a loop that is reading keys for the current Nvim
mode, which will be handled by calling event_process()
This approach has the advantage of integrating nicely with the current
codebase, eg: vimscript code can be executed asynchronously with little
surprises(Its the same as if the user typed a key).
The problem with using keys to represent any event is that it also interferes with
operators, and not every event needs or should do that. For example, consider
this scenario:
- A msgpack-rpc client calls vim_feedkeys("d")
- Nvim processes K_EVENT, pushing "d" to the input queue
- Nvim processes "d", entering operator-pending mode to wait for a motion
- The client calls vim_feedkeys("w"), expecting Nvim to delete a word
- Nvim processes K_EVENT, breaking out of operator-pending and pushing "w"
- Nvim processes "w", moving a word
This commit fixes the above problem by removing all automatic calls to
`event_push`(which is what generates K_EVENT input). Right now this also breaks
redrawing initiated by asynchronous events(and possibly other stuff too, Nvim is
a complex state machine and we can't simply run vimscript code anywhere).
In future commits the calls to `event_push` will be inserted only where it's
absolutely necessary to run code in "key reading loops", such as when executing
vimscript code or mutating editor data structures in ways that currently can
only be done by the user.
This approach is more flexible because we don't need to support a fixed set of
"event types", any module can push events to be handled in main loop by simply
passing a callback to the Event structure.
rbuffer_data was renamed to rbuffer_read_ptr, and it represents the next read
position in a RBuffer instance. Similarly, rbuffer_write_ptr was added to
represent the next write position.
Also, rbuffer_data was being used for writing(in alloc_cb), replace that by
rbuffer_write_ptr.
RBuffer instances represent the internal buffer used by RStreams.
This changes RStream constructor to receive RBuffer pointers and adds a set of
RBuffer methods that expose the lower level buffer manipulation to consumers of
the RStream API.
If we fail to bind to the server address, do not try and listen lest the
reported error always be "invalid argument".
Also, return whether or not we errored from server_init() in case we
want to respond differently in the future.
The os_system function uses a write callback to close the input stream when the
write completes, but this causes a memory error because the callback is invoked
right before the stream is freed by the caller.
This fixes the problem by removing the callback set by os_system. Instead, it
calls job_close_in immediately after writing(the stream will only close after
the write completes). The 'pending' parameter was also removed from the
'write_cb' as it should be hidden by the wstream module.
While the `wstream_set_write_cb` and `job_write_cb` are no longer used, they
will remain in the codebase for future use.
