libuv offers considerable child process management, abstracting the platform differences and allowing communication with the child process using streams or named pipes.
A common idiom in Unix is for every process to do one thing and do it well. In such a case, a process often uses multiple child processes to achieve tasks (similar to using pipes in shells). A multi-process model with messages may also be easier to reason about compared to one with threads and shared memory.
A common refrain against event-based programs is that they cannot take advantage of multiple cores in modern computers. In a multi-threaded program the kernel can perform scheduling and assign different threads to different cores, improving performance. But an event loop has only one thread. The workaround can be to launch multiple processes instead, with each process running an event loop, and each process getting assigned to a separate CPU core.
The simplest case is when you simply want to launch a process and know when it
exits. This is achieved using uv_spawn.
spawn/main.c
.. literalinclude:: ../code/spawn/main.c
:linenos:
:lines: 6-8,15-
:emphasize-lines: 11,13-17
Note
options is implicitly initialized with zeros since it is a global
variable. If you change options to a local variable, remember to
initialize it to null out all unused fields:
uv_process_options_t options = {0};
The uv_process_t struct only acts as the handle, all options are set via
uv_process_options_t. To simply launch a process, you need to set only the
file and args fields. file is the program to execute. Since
uv_spawn uses execvp internally, there is no need to supply the full
path. Finally as per underlying conventions, the arguments array has to be
one larger than the number of arguments, with the last element being NULL.
After the call to uv_spawn, uv_process_t.pid will contain the process
ID of the child process.
The exit callback will be invoked with the exit status and the type of signal which caused the exit.
spawn/main.c
.. literalinclude:: ../code/spawn/main.c
:linenos:
:lines: 9-12
:emphasize-lines: 3
It is required to close the process watcher after the process exits.
Before the child process is launched you can control the execution environment
using fields in uv_process_options_t.
Set uv_process_options_t.cwd to the corresponding directory.
uv_process_options_t.env is a null-terminated array of strings, each of the
form VAR=VALUE used to set up the environment variables for the process. Set
this to NULL to inherit the environment from the parent (this) process.
Setting uv_process_options_t.flags to a bitwise OR of the following flags,
modifies the child process behaviour:
UV_PROCESS_SETUID- sets the child's execution user ID touv_process_options_t.uid.UV_PROCESS_SETGID- sets the child's execution group ID touv_process_options_t.gid.
Changing the UID/GID is only supported on Unix, uv_spawn will fail on
Windows with UV_ENOTSUP.
UV_PROCESS_WINDOWS_VERBATIM_ARGUMENTS- No quoting or escaping ofuv_process_options_t.argsis done on Windows. Ignored on Unix.UV_PROCESS_DETACHED- Starts the child process in a new session, which will keep running after the parent process exits. See example below.
Passing the flag UV_PROCESS_DETACHED can be used to launch daemons, or
child processes which are independent of the parent so that the parent exiting
does not affect it.
detach/main.c
.. literalinclude:: ../code/detach/main.c
:linenos:
:lines: 9-30
:emphasize-lines: 12,19
Just remember that the handle is still monitoring the child, so your program
won't exit. Use uv_unref() if you want to be more fire-and-forget.
libuv wraps the standard kill(2) system call on Unix and implements one
with similar semantics on Windows, with one caveat: all of SIGTERM,
SIGINT and SIGKILL, lead to termination of the process. The signature
of uv_kill is:
uv_err_t uv_kill(int pid, int signum);
For processes started using libuv, you may use uv_process_kill instead,
which accepts the uv_process_t watcher as the first argument, rather than
the pid. In this case, remember to call uv_close on the watcher.
libuv provides wrappers around Unix signals with some Windows support as well.
Use uv_signal_init() to initialize
a handle and associate it with a loop. To listen for particular signals on
that handler, use uv_signal_start() with the handler function. Each handler
can only be associated with one signal number, with subsequent calls to
uv_signal_start() overwriting earlier associations. Use uv_signal_stop() to
stop watching. Here is a small example demonstrating the various possibilities:
signal/main.c
.. literalinclude:: ../code/signal/main.c
:linenos:
:emphasize-lines: 17-18,27-28
Note
uv_run(loop, UV_RUN_NOWAIT) is similar to uv_run(loop, UV_RUN_ONCE)
in that it will process only one event. UV_RUN_ONCE blocks if there are no
pending events, while UV_RUN_NOWAIT will return immediately. We use NOWAIT
so that one of the loops isn't starved because the other one has no pending
activity.
Send SIGUSR1 to the process, and you'll find the handler being invoked
4 times, one for each uv_signal_t. The handler just stops each handle,
so that the program exits. This sort of dispatch to all handlers is very
useful. A server using multiple event loops could ensure that all data was
safely saved before termination, simply by every loop adding a watcher for
SIGINT.
A normal, newly spawned process has its own set of file descriptors, with 0,
1 and 2 being stdin, stdout and stderr respectively. Sometimes you
may want to share file descriptors with the child. For example, perhaps your
applications launches a sub-command and you want any errors to go in the log
file, but ignore stdout. For this you'd like to have stderr of the
child be the same as the stderr of the parent. In this case, libuv supports
inheriting file descriptors. In this sample, we invoke the test program,
which is:
proc-streams/test.c
.. literalinclude:: ../code/proc-streams/test.c
The actual program proc-streams runs this while sharing only stderr.
The file descriptors of the child process are set using the stdio field in
uv_process_options_t. First set the stdio_count field to the number of
file descriptors being set. uv_process_options_t.stdio is an array of
uv_stdio_container_t, which is:
.. literalinclude:: ../libuv/include/uv.h
:lines: 826-834
where flags can have several values. Use UV_IGNORE if it isn't going to be
used. If the first three stdio fields are marked as UV_IGNORE they'll
redirect to /dev/null.
Since we want to pass on an existing descriptor, we'll use UV_INHERIT_FD.
Then we set the fd to stderr.
proc-streams/main.c
.. literalinclude:: ../code/proc-streams/main.c
:linenos:
:lines: 15-17,27-
:emphasize-lines: 6,10,11,12
If you run proc-stream you'll see that only the line "This is stderr" will
be displayed. Try marking stdout as being inherited and see the output.
It is dead simple to apply this redirection to streams. By setting flags
to UV_INHERIT_STREAM and setting data.stream to the stream in the
parent process, the child process can treat that stream as standard I/O. This
can be used to implement something like CGI.
A sample CGI script/executable is:
cgi/tick.c
.. literalinclude:: ../code/cgi/tick.c
The CGI server combines the concepts from this chapter and :doc:`networking` so that every client is sent ten ticks after which that connection is closed.
cgi/main.c
.. literalinclude:: ../code/cgi/main.c
:linenos:
:lines: 49-63
:emphasize-lines: 10
Here we simply accept the TCP connection and pass on the socket (stream) to
invoke_cgi_script.
cgi/main.c
.. literalinclude:: ../code/cgi/main.c
:linenos:
:lines: 16, 25-45
:emphasize-lines: 8-9,18,20
The stdout of the CGI script is set to the socket so that whatever our tick
script prints, gets sent to the client. By using processes, we can offload the
read/write buffering to the operating system, so in terms of convenience this
is great. Just be warned that creating processes is a costly task.
libuv's uv_pipe_t structure is slightly confusing to Unix programmers,
because it immediately conjures up | and pipe(7). But uv_pipe_t is
not related to anonymous pipes, rather it is an IPC mechanism. uv_pipe_t
can be backed by a Unix Domain Socket or Windows Named Pipe to allow
multiple processes to communicate. This is discussed below.
A parent and child can have one or two way communication over a pipe created by
settings uv_stdio_container_t.flags to a bit-wise combination of
UV_CREATE_PIPE and UV_READABLE_PIPE or UV_WRITABLE_PIPE. The
read/write flag is from the perspective of the child process.
Since domain sockets [1] can have a well known name and a location in the file-system they can be used for IPC between unrelated processes. The D-BUS system used by open source desktop environments uses domain sockets for event notification. Various applications can then react when a contact comes online or new hardware is detected. The MySQL server also runs a domain socket on which clients can interact with it.
When using domain sockets, a client-server pattern is usually followed with the creator/owner of the socket acting as the server. After the initial setup, messaging is no different from TCP, so we'll re-use the echo server example.
pipe-echo-server/main.c
.. literalinclude:: ../code/pipe-echo-server/main.c
:linenos:
:lines: 70-
:emphasize-lines: 5,10,14
We name the socket echo.sock which means it will be created in the local
directory. This socket now behaves no different from TCP sockets as far as
the stream API is concerned. You can test this server using socat:
$ socat - /path/to/socket
A client which wants to connect to a domain socket will use:
void uv_pipe_connect(uv_connect_t *req, uv_pipe_t *handle, const char *name, uv_connect_cb cb);
where name will be echo.sock or similar.
The cool thing about domain sockets is that file descriptors can be exchanged between processes by sending them over a domain socket. This allows processes to hand off their I/O to other processes. Applications include load-balancing servers, worker processes and other ways to make optimum use of CPU. libuv only supports sending TCP sockets or other pipes over pipes for now.
To demonstrate, we will look at a echo server implementation that hands of clients to worker processes in a round-robin fashion. This program is a bit involved, and while only snippets are included in the book, it is recommended to read the full code to really understand it.
The worker process is quite simple, since the file-descriptor is handed over to it by the master.
multi-echo-server/worker.c
.. literalinclude:: ../code/multi-echo-server/worker.c
:linenos:
:lines: 7-9,81-
:emphasize-lines: 6-8
queue is the pipe connected to the master process on the other end, along
which new file descriptors get sent. It is important to set the ipc
argument of uv_pipe_init to 1 to indicate this pipe will be used for
inter-process communication! Since the master will write the file handle to the
standard input of the worker, we connect the pipe to stdin using
uv_pipe_open.
multi-echo-server/worker.c
.. literalinclude:: ../code/multi-echo-server/worker.c
:linenos:
:lines: 51-79
:emphasize-lines: 10,15,20
First we call uv_pipe_pending_count() to ensure that a handle is available
to read out. If your program could deal with different types of handles,
uv_pipe_pending_type() can be used to determine the type.
Although accept seems odd in this code, it actually makes sense. What
accept traditionally does is get a file descriptor (the client) from
another file descriptor (The listening socket). Which is exactly what we do
here. Fetch the file descriptor (client) from queue. From this point
the worker does standard echo server stuff.
Turning now to the master, let's take a look at how the workers are launched to allow load balancing.
multi-echo-server/main.c
.. literalinclude:: ../code/multi-echo-server/main.c
:linenos:
:lines: 9-13
The child_worker structure wraps the process, and the pipe between the
master and the individual process.
multi-echo-server/main.c
.. literalinclude:: ../code/multi-echo-server/main.c
:linenos:
:lines: 51,61-95
:emphasize-lines: 17,20-21
In setting up the workers, we use the nifty libuv function uv_cpu_info to
get the number of CPUs so we can launch an equal number of workers. Again it is
important to initialize the pipe acting as the IPC channel with the third
argument as 1. We then indicate that the child process' stdin is to be
a readable pipe (from the point of view of the child). Everything is
straightforward till here. The workers are launched and waiting for file
descriptors to be written to their standard input.
It is in on_new_connection (the TCP infrastructure is initialized in
main()), that we accept the client socket and pass it along to the next
worker in the round-robin.
multi-echo-server/main.c
.. literalinclude:: ../code/multi-echo-server/main.c
:linenos:
:lines: 31-49
:emphasize-lines: 9,12-13
The uv_write2 call handles all the abstraction and it is simply a matter of
passing in the handle (client) as the right argument. With this our
multi-process echo server is operational.
Thanks to Kyle for pointing out that uv_write2() requires a non-empty
buffer even when sending handles.
| [1] | In this section domain sockets stands in for named pipes on Windows as well. |