Re: [PATCH] Improve LinuxThreads compatibility in rfork()

On Mon, Jul 11, 2011 at 04:23:36PM +0200, Petr Salinger wrote:
> >>>Can you, please, describe the reasoning behind the
> >>>>+	        if (sig == SIGCHLD) sig = 0;
> >>>line ?
> >>
> >>The main reason is backward compatibility.
> >>The original FreeBSD code allows only to select between
> >>SIGUSR1 or SIGCHLD signals.
> >>
> >>The our extension changes meaning of RFLINUXTHPN to select signal based on
> >>bits 24-30 of passed flags instead of using SIGUSR1 every time.
> >>
> >>When the passed "signal" number is zero, it should behave identically
> >>on plain FreeBSD and in our environment, therefore SIGUSR1 is selected.
> >>The assumption is (have been) that (yet) undefined bits are zero.
> >>That way we are backward compatible with original FreeBSD.
> >>
> >>We still need an alternative way to select "none signal is sent"
> >>after child exit (under linux #0 is used).
> >>
> >>The SIGCHLD can be "selected" (also on original FreeBSD) by not specifying
> >>RFLINUXTHPN, therefore combination of RFLINUXTHPN and passed "signal"
> >>number SIGCHLD is (have been) used for "none signal".
> >>
> >>BTW, the opposite side is in
> >>
> >>http://anonscm.debian.org/viewvc/glibc-bsd/trunk/glibc-ports/kfreebsd/clone.c?view=markup
> >
> >I shall state that the sig == SIGCHLD case is ugly. Having the separate
> >flag "do not send signal to the parent" would be much less clumsy.
> >What are the requirements for the ABI stability for Debian/kFreeBSD ?
> >Can this be fixed now, or is it too late ?
> 
> It should be backward compatible with one previous version.
> 
> What about in long term this:
> 
> RFLINUXTHPN bit will be renamed and will have meaning
> "select signal based on bits 24-30 of passed flags"
> 
> - zero would mean "no signal"
> - SIGCHLD would mean undefined
> - SIGUSR1 would mean SIGUSR1
> 
> It is ABI/API breakage under original FreeBSD.
> The question is how frequently RFLINUXTHPN is used under native FreeBSD
> and its port collection.
RFLINUXPTH was used by the linuxthreads port, that was popular in the
time of FreeBSD 4.x and may be 5.x to run mysql. I will object against
this breakage.

> 
> And under "Debian GNU/kFreeBSD COMPAT" or 8-COMPAT
> - SIGCHLD would mean "no signal"
We never user COMPAT to _change_ the meaning of something, only to
exclude some functionality, like syscall.

> 
> We do not use SIGUSR1 currently, the eglibc side can detect whether
> it runs under new-enough kernel and decide whether use 0 or SIGCHLD
> for "no signal".
> 
> The kernel side would be something like:
> 
>     if (flags & RFLINUXTHPN)
>     {
>         p2->p_sigparent = RFTHPNSIGNUM(flags);
> #if COMPAT8
>         if (p2->p_sigparent == SIGCHLD)
>            p2->p_sigparent = 0;
> #endif
>     }
No, this is even uglier, and see the note about compat.

BTW, it seems that our rfork(2) ignores unknown flags. This should be
fixed, might be in the same patch.