The process_t pointer sent to setup_child_process can actually be 0
without it being failure, as that is what fish sends when `exec` is run
(in the case of INTERNAL_EXEC).
This was causing exec to fail.
There is no more race condition between parent and child with
regards to setting the process groups. Each child sets it for themselves
and then blocks indefinitely until the parent does what it needs to for
them (having waited for them to set their process groups). They are not
SIGCONT'd until the next process in the chain (if any) starts so that
that process can join their process group and open the pipes.
In the last commit, we introduced an indiscriminate if !EXTERNAL check
that unblocks a previously SIGSTOP'd command (if any) to allow the main
loop in exec_job to read from it without deadlocking (since builtins and
functions read directly from input as an optimization, sometimes).
Now only unblocking where a fork will not happen to ensure that if a
builtin ends up forking, that fork'd process is guaranteed to be able to
join the previous process' process group and access its output pipes.
Setting the process group in a fork/exec scenario is a well-documented
race condition in pretty much any job control mechanism [0] [1]. The
Wikipedia article contradicts the glibc article and suggests that the
best approach is for the parent to wait for the child to become the
process group leader, while the glibc article suggests that both should
make it so (which is what fish did previously). However, I'm running
into cases where tcsetpgrp is causing an EPERM error, which it isn't
documented to do except if the session id for the calling process
differs from that of the target process group (which is never the case
in fish since they are all part of the same session), which should cause
a _different_ error (SIGTTOU to be sent to all members of the calling
process' group).
In all cases, this is easily remedied by checking if the process group
in question is already in control of the terimnal. There's still the
off-chance that in the time between we check that and the time that the
command completes that situation may have changed, but the parent
process is supposed to ignore the result of this call if it errors out.
[0]: https://en.wikipedia.org/wiki/Process_group
[1]: https://www.gnu.org/software/libc/manual/html_node/Launching-Jobs.html
We were having child processes SIGSTOP themselves immediately after
setting their process group and before launching their intended targets,
but they were not necessarily stopped by the time the next command was
being executed (so the opposite of the original race condition where
they might have finished executing by the time the next command came
around), and as a result when we sent them SIGCONT, that could never
reach. Now using waitpid to synchronize the SIGSTOP/SIGCONT between the
two.
If we had a good, unnamed inter-process event/semaphore, we could use
that to have a child process conditionally stop itself if the next
command in the job chain hadn't yet been started / setup, but this is
probably a lot more straightforward and less-confusing, which isn't a
bad thing.
Additionally, there was a bug caused by the fact that the main exec_job
loop actually blocks to read from previous commands in the job if the
current command is a built-in that doesn't need to fork.
With this waitpid code, I was able to finally add the SIGSTOP code to
all the fork'd processes in the main exec_job loop without introducing
deadlocks; it turns out that they should be treated just like the main
EXTERNAL fork, but they tend to execute faster causing the same deadlock
described above to occur more readily.
The only thing I'm not sure about is whether we should execute
unblock_pid undconditionally for all !EXTERNAL commands. It makes more
sense to *only* do that if a blocking read were about to be done in the
main loop, otherwise the original race condition could still appear
(though it is probably mitigated by whatever duration the SIGSTOP lasted
for, even if it is SIGCONT'd before the next command tries to join the
process group).
I hadn't realized that the for loop is called multiple times for a given
"single input" (anything that doesn't include semicolons, etc) to fish,
and so processes were being blocked but blocked_pid was lost by the time
that the next job (which was reading from the last process in the
previous job) came around.
Now using a static variable to store the last blocked PID. AFAICT, this
main job control loop is always executed from the same process and
thread, so this shouldn't need to be wrapped in atomics/mutexes, etc.
This code should be more portable, and certainly cleaner. We are
currently always sending SIGCONT to the last process (if it was part of
a job chain) regardless of whether it called SIGSTOP on itself or not,
which should be fine.
Need to explore whether or not the other forks in src/exec.cpp need to
be SIGSTOP'd on run or only the one that we included in this patch.
I'm not sure if this happens on all platforms, but under WSL with the
existing codebase, processes in the job chain that pipe their
stdout/stderr to the next process in the job could terminate before the
next job started (on fast enough machines for quick enough jobs).
This caused issues like #4235 and possibly #3952, at least for external
commands. What was happening is that the first process was finishing
before the second process was fully set up. fish would then try to
assign (in both the child and the parent) the process group id belonging
to the process group leader to the new process; and if the first process
had already terminated, it would have ended its process group with it as
well before that happened.
I'm not sure if there was already a mechanism in place for ensuring that
a process remains running at least as long as it takes for the next
process in the chain to join its group, etc., but if that code was
there, it wasn't working in my test setup (WSL).
This patch definitely needs some review; I'm not sure how I should
handle non-external commands (and external commands executed via
posix_spawn). I don't know if they are affected by the race condition in
the first place, but when I tried to add the same "wait for next command
in chain to run before unblocking" that would cause black screens
requiring ctrl+c to bypass.
The "unblock previous command" code was originally run by the next child
to be forked, but was then moved to the shell code instead, making it
more-centrally located and less error-prone.
Note that additional headers may be required for the mmap system call on
other platforms.
This is the first step to implementing issue #4200 is to stop subclassing
env_var_t from wcstring. Not too surprisingly doing this identified
several places that were incorrectly treating env_var_t and wcstring as
interchangeable types. I'm not talking about those places that passed
an env_var_t instance to a function that takes a wcstring. I'm talking
about doing things like assigning the former to the latter type, relying
on the implicit conversion, and thus losing information.
We also rename `env_get_string()` to `env_get()` for symmetry with
`env_set()` and to make it clear the function does not return a string.