system() using vfork() or posix_spawn()

[Konstantin Belousov]

On Sun, Aug 05, 2012 at 11:54:32PM +0200, Jilles Tjoelker wrote:
> On Mon, Jul 30, 2012 at 01:53:03PM +0300, Konstantin Belousov wrote:
> > On Mon, Jul 30, 2012 at 12:24:08PM +0200, Jilles Tjoelker wrote:
> > > People sometimes use system() from large address spaces where it would
> > > improve performance greatly to use vfork() instead of fork().
> 
> > > A simple approach is to change fork() to vfork(), although I have not
> > > tried this. It seems safe enough to use sigaction and sigprocmask system
> > > calls in the vforked process.
> 
> > > Alternatively, we can have posix_spawn() do the vfork() with signal
> > > changes. This avoids possible whining from compilers and static
> > > analyzers about using vfork() in system.c. However, I do not like the
> > > tricky code for signals and that it adds lines of code.
> 
> > > This is lightly tested.
> 
> > It is interesting to note that for some time our vfork(2) no longer
> > stops the whole forked process (parent), only the forking thread is
> > waiting for the child exit or exec. I am not sure is this point
> > important for system(3), but determined code can notice the difference
> > from the fork->vfork switch.
> 
> Neither fork nor vfork call thread_single(SINGLE_BOUNDARY), so this is
> not a difference.
It is the difference, because vforked child shares parent address space.

> 
> Thread singling may be noticeable from a failing execve() (but only in
> the process doing execve()) and in the rare case of rfork() without
> RFPROC.

No, other running threads in parent affect vforked child till exec or exit.
In fact, I would classify this as bug, but not a serious one.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 196 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-hackers/attachments/20120806/f44c18de/attachment.pgp