[PATCH v2 22/46] vfork: Use fork if arch does not have the vfork syscall

[Rich Felker]

On Tue, Nov 27, 2012 at 02:09:36PM +0530, Vineet Gupta wrote:
> > diff --git a/libc/sysdeps/linux/common/vfork.c b/libc/sysdeps/linux/common/vfork.c
> > index e7c9208..377418c 100644
> > --- a/libc/sysdeps/linux/common/vfork.c
> > +++ b/libc/sysdeps/linux/common/vfork.c
> > @@ -4,13 +4,27 @@
> >   * Licensed under the LGPL v2.1, see the file COPYING.LIB in this tarball.
> >   */
> >  
> > +#include <signal.h>
> >  #include <unistd.h>
> >  #include <sys/types.h>
> >  #include <sys/syscall.h>
> >  
> >  extern __typeof(vfork) __vfork attribute_hidden;
> >  
> > -#ifdef __NR_vfork
> > +#if defined(__NR_fork) && !defined(__NR_vfork)
> > +pid_t __vfork(void)
> > +{
> > +	pid_t pid = INLINE_SYSCALL(fork, 0);
> > +
> > +	if (pid<0)
> > +		return -1
> > +
> > +	return pid;
> > +}
> > +weak_alias(__vfork, vfork)
> > +libc_hidden_weak(vfork)
> > +
> > +#elif defined(__NR_vfork)
> >  
> >  # define __NR___vfork __NR_vfork
> >  _syscall0(pid_t, __vfork)
> > 
>  [...]
> While v1 of this particular patch made sense to me - this one doesn't.
> kernel's asm-generic unistd.h only defines clone (and doesn't define
> vfork/fork). Thus this file will compile to nothing as opposed to v1
> where vfork would correctly call clone syscall with correct args -
> something which the vfork syscall handler itself in kernel typically
> does. And since that would be gone - we need to do same thing in user
> space wrapper.

The new patch is wrong too. The correct way to do vfork from C on a
no-legacy-syscalls kernel is to use the clone syscall with the flags
combination that makes it equivalent to plain fork (i.e. just
SIGCHLD). Using CLONE_VFORK from C is incorrect because there's no way
to preserve the return address for the parent to use after the child
runs.

Rich