[boost] Boost.Process article: Tutorial and request for comments

Subject: [boost] Boost.Process article: Tutorial and request for comments
From: Boris Schaeling (boris_at_[hidden])
Date: 2009-04-21 08:01:09

I've written an article about Boost.Process:
http://www.highscore.de/cpp/process/

I've also fixed a few bugs in Boost.Process (which is now 0.31):
http://www.highscore.de/boost/process.zip

And the latest source code can also be found in the sandbox:
https://svn.boost.org/svn/boost/sandbox/process/

For developers not familar with Boost.Process the article is a tutorial to
get more easily started. For developers familiar with Boost.Process it is
a request for comments.

I've been looking at the big picture finally trying to find out where and
how Boost.Process should be changed and in which direction it should move
to become a complete library one day. As not everyone will read the entire
article here are my thoughts:

1) Boost.Process is based on a few concepts like 'process' and 'context'.
While those make sense there are other concepts where I wonder if they
should be dropped:

* 'executable' is the executable name which in practice must be a
std::string. I think the original idea was to support numerous string
types. This leads to the question how to support different Unicode
encodings which I think is still unanswered in the Boost community. Other
libraries, eg. Boost.Interprocess, simply stick to one string type
(typically std::string).

* 'handle' is the underlying native file handle of a stream (file
descriptor on POSIX systems, handle on Windows). The concept is based on a
class in the detail namespace of Boost.Process. There is just one class
and it's not meant to be used by users of the library either. Thus I'm not
sure what's the benefit of having an explicit 'handle' concept at all.

2) Environment variables must be sorted on Windows but not on POSIX
systems. On Windows the type environment could be a typedef for std::map
and on POSIX systems for std::unordered_map. Currently it's always a
typedef for std::map. Is that ok? Should it be changed? Should a macro be
provided in case developers want to use different typedefs?

3) Currently Boost.Process is really only useful for creating child
processes. It's not possible for example to detect and iterate over other
processes on a system. I guess Boost.Process should be enhanced here. Then
one day the well-known utility 'ps' on POSIX systems could be implemented
in Boost C++. Any comments?

4) When a process is created the standard streams (stdin, stdout and
stderr) are closed by default. Is this a good choice? If it isn't what
should happen otherwise? Inherting standard streams (which is the default
behavior with fork on POSIX systems)?

5) Boost.Process requires to specify an absolute path to an executable
which should be started. There is a helper function which returns the
absolute path for an executable name (it uses the environment variable
PATH). The question is if this should be done automatically if no absolute
path is given?

6) When a stream of a child process is captured Boost.Process
automatically creates a pipe. As pipes don't support asynchronous
operations on Windows a named pipe is required (called FIFO on POSIX
systems). Currently a macro must be defined to make Boost.Process use a
named pipe on Windows. It might make more sense to enable developers to
choose either a pipe or a FIFO explicitly - not only on Windows but on
all platforms. Any comments?

7) It's possible to capture and access streams of a child process. Those
streams are derived from std::istream and std::ostream and thus only
support blocking I/O operations. Currently the underlying handle has to be
fetched to initialize a generic I/O object from Boost.Asio to use
asynchronous I/O operations. How should the library be changed to make it
easier to use asynchronous I/O?

8) There is a method wait() which makes it possible to wait for another
process to exit. Currently wait() is provided by the class child though.
Shouldn't it really be defined by the class process (so you can wait not
only for child processes to exit)?

9) As wait() blocks an asynchronous alternative should be provided.

10) The Windows-specific classes all start with win32_ (eg. win32_child).
Wouldn't it make more sense to use windows_ instead (like windows_child)
or a windows namespace? First other libraries use windows_ (see
Boost.Interprocess). Second there should be no reason why a class like
win32_child can't be used on a 64-bit Windows.

Boris