What is vfork()?

[chris at umcp-cs.UUCP]

Vfork() is intended to be a more efficient version of fork(),
especially for large programs.  It is mainly useful for programs
wishing to exec() some other program.  The idea behind vfork() is
to not actually copy all the pages of the program, when it isn't
going to use most of them anyway.

4.1 behavior:

	Like fork(), vfork() produces a copy of a process.  Unlike
	fork(), it doesn't actually copy the process.  Instead it
	fiddles with kernel data structures a bit, copies the page
	tables (memory allocation list) for the process to the new
	process, and suspends the "parent" (who has, at the moment,
	no memory, since it's all given to the "child").  The child
	gets to run until it exit()s or exec()s something else.
	Then, instead of releasing the memory/page tables/etc, the
	kernel gives everything back to the parent and lets it
	continue.

	This "cheating" has the interesting side effect that the
	code:

	main () {
		int i = 0;

		if (vfork() == 0)	/* child */
			exit (i = 10);
		printf ("i = %d\n", i);
	}

	prints "10".  The child can modify the parent!  However, this
	is not useful for IPC since the parent is held suspended until
	the child exit()s or exec()s.  Also, this behavior is considered
	wrong, and will change (has changed?) in future releases of
	Berkeley Unix.  Oh yeah, if "i" were a "register int" you'd
	get 0 (howzat for consistency?).

4.2:	(maybe, maybe not yet, ``definitely'' in 4.3)
	The vfork() system call will be (has been?) changed to allow
	both processes to run simultaneously, sharing memory resources
	as long as the child does not attempt to write to the shared
	memory.  If the child does write to the shared memory, the
	write is trapped and the memory is copied to a new page (or,
	if necessary, depending on hardware, maybe the entire process
	is copied) and the child shares one less page (or no pages).
	This will actually be done to the fork() system call as the
	behavior of fork() and vfork() will then be identical.  The
	new implementation is merely much more efficient (on paged
	systems anyway).
-- 
In-Real-Life: Chris Torek, Univ of MD Comp Sci
UUCP:	{seismo,allegra,brl-bmd}!umcp-cs!chris
CSNet:	chris at umcp-cs		ARPA:	chris.umcp-cs at CSNet-Relay