This concerns builtins writing to an io_buffer_t. io_buffer_t is how fish
captures output, especially in command substitutions:
set STUFF (string upper stuff)
Recall that io_buffer_t fills itself by reading from an fd (typically
connected to stdout of the command). However if our command is a builtin,
then we can write to the buffer directly.
Prior to this change, when a builtin anticipated writing to an
io_buffer_t, it would first write into an internal buffer, and then after
the builtin was finished, we would copy it to the io_buffer_t. This was
because we didn't have a polymorphic receiver for builtin output: we
always buffered it and then directed it to the io_buffer_t or file
descriptor or stdout or whatever.
Now that we have polymorphpic io_streams_t, we can notice ahead of time
that the builtin output is destined for an internal buffer and have it
just write directly to that buffer. This saves a buffering step, which is
a nice simplification.
io_buffer_t is used to buffer output from a command substitution, so we
can split it into arguments. Typically io_buffer_t reads from its pipe
until it gets EOF and then stops reading. However it may be that the
cmdsub ends but EOF is not delivered because the stdout of the cmdsub
escaped with a background process.
Prior to this change we would wake up every 100 msec (select timeout) to
check if the cmdsub is finished. However this 100 msec adds latency if a
background process is launched from e.g. fish_prompt.
Switch to the new poke() function. Now when the cmdsub is finished, it
pokes its item, which explicitly wakes it up. This removes the extra
latency.
Fixes#7559
It was possible though unlikely for make_autoclose_pipes to close only
one side of pipe, if it fails to find a new fd. This would result in an
fd leak. Ensure that doesn't happen.
In principle this would allow 'string split' or whatever to output to
stderr and not lose the item separation. In practice this is not used
but it fixes a TODO.
builtins output to stdout and stderr via io_streams_t. Prior to this fix, it
contained an output_stream_t which just wraps a buffer. So all builtin output
went to this buffer (except for eval).
Switch output_stream_t to become a new abstract class which can output to a
buffer, file descriptor, or nowhere. This allows for example `string` to stream
its output as it is produced, instead of buffering it.
Initially I wanted to pick a different name to avoid confusion with
process groups, but really job trees *are* process groups. So name them
to reflect that fact.
Also rename "placeholder" to "internal" which is clearer.
Job trees come in two flavors: “placeholders” for jobs which are only fish
functions, and non-placeholders which need to track a pgid. This adds
logic to allow a job to decide if its parent's job tree is appropriate,
and allocating a new tree if not.
Currently fish aborts execution mid-pipeline if a file redirection
failed, which can leave the shell in a broken state (job abandoned after
giving control of the terminal to an already-executed job in the
pipeline).
This patch replaces a failed fd with a closed fd and continues execution
if the affected process wasn't the first in the pipeline.
While this is a hack to address the regression behind fish-shell/#7038
introduced in d62576c, it can also be argued that this behavior is
actually more correct... right?
Closes#7038.
builtin_eval needs to know whether to set up bufferfills to capture its
output and/or errput; it should do this specifically if the output and
errput is piped (and not, say, directed to a file). In preparation for
this change, add bools to io_streams_t which track whether stdout and
stderr are specifically piped.
Prior to this fix, builtin_eval would direct output to the io_chain of the
job. The problem is with pipes: `builtin_eval` might happily attempt to
write unlimited output to the write end of a pipe, but the corresponding
reading process has not yet been launched. This results in deadlock.
The fix is to buffer all the output from `builtin_eval`. This is not fun
but the best that can be done until we have real concurrent processes.
Fixes#6806
This switches bufferfills from using an exclusively-owned thread, to
sharing an fd_monitor. This allows multiple bufferfills to all use the same
thread.
user_supplied was used to distinguish IO redirections which were
explicit, vs those that came about through "transmogrphication." But
transmogrification is no more. Remove the flag.
Prior to this fix, a file redirection was turned into an io_file_t. This is
annoying because every place where we want to apply the redirection, we
might fail due to open() failing. Switch to opening the file at the point
we resolve the redirection spec. This will simplify a lot of code.
Prior to this change, a process after it has been constructed by
parse_execution, but before it is executed, was given a list of
io_data_t redirections. The problem is that redirections have a
sensitive ownership policy because they hold onto fds. This made it
rather hard to reason about fd lifetime.
Change these to redirection_spec_t. This is a textual description
of a redirection after expansion. It does not represent an open file and
so its lifetime is no longer important.
This enables files to be held only on the stack, and are no longer owned
by a process of indeterminate lifetime.
fish has to ensure that the pipes it creates do not conflict with any
explicit fds named in redirections. Switch this code to using
autoclose_fd_t to make the ownership logic more explicit, and also
introduce fd_set_t to reduce the dependence on io_chain_t.
Background fillthreads are used when we want to populate a buffer from an
external command. The most common is command substitution.
Prior to this commit, fish would spin up a fillthread whenever required.
This ended up being quite expensive.
Switch to using the iothread pool instead. This enables reusing the same
thread(s), which prevents needing to spawn new threads. This shows a big
perf win on the alias benchmark (766 -> 378 ms).
Previously when propagating explicitly separated output, we would early-out
if the buffer was empty, where empty meant contains no characters. However
it may contain one or more empty strings, in which case we should propagate
those strings.
Remove this footgun "empty" function and handle this properly.
Fixes#5987
This runs build_tools/style.fish, which runs clang-format on C++, fish_indent on fish and (new) black on python.
If anything is wrong with the formatting, we should fix the tools, but automated formatting is worth it.
In a job, a deferred process is the last fish internal process which pipes
to an external command. Execute the deferred process last; this will allow
for streaming its output.
Now that we use an internal process to perform builtin output, simplify the
logic around how it is performed. In particular we no longer have to be
careful about async-safe functions since we do not fork.
Also fix a bunch of comments that no longer apply.
This is a large change to how io_buffers are filled. The essential problem
comes about with code like (example):
echo ( /bin/pwd )
The output of /bin/pwd must go to fish, not the tty. To arrange for this,
fish does the following:
1. Invoke pipe() to create a pipe.
2. Add an io_bufferfill_t redirection that owns the write end of the pipe.
3. After fork (or equiv), call dup2() to replace pwd's stdout with this pipe.
Now when /bin/pwd writes, it will send output to the read end of the pipe.
But who reads it?
Prior to this fix, fish would do the following in a loop:
1. select() on the pipe with a 10 msec timeout
2. waitpid(WNOHANG) on the pwd proc
This polling is ugly and confusing and is what is replaced here.
With this new change, fish now reads from the pipe via a background thread:
1. Spawn a background pthread, which select()s on the pipe's read end with
a long (100 msec) timeout.
2. In the foreground, waitpid() (allowing hanging) on the pwd proc.
The big win here is a major simplification of job_t::continue_job() since
it no longer has to worry about filling buffers. This will make things
easier for concurrent execution.
It may not be obvious why the background thread still needs a poll (100 msec).
The answer is for cases where the write end of the fd escapes, in particular
background processes invoked inside command substitutions. psub is perhaps
the only important case of this (other shells typically just hang here).
This makes some significant architectual improvements to io_pipe_t and
io_buffer_t.
Prior to this fix, io_buffer_t subclassed io_pipe_t. io_buffer_t is now
replaced with a class io_bufferfill_t, which does not subclass pipe.
io_pipe_t no longer remembers both fds. Instead it has an autoclose_fd_t,
so that the file descriptor ownership is clear.
This represents a "resolved" io_chain_t, where all of the different io_data_t
types have been reduced to a sequence of dup2() and close(). This will
eliminate a lot of the logic duplication around posix_spawn vs fork, and pave
the way for in-process redirections.
