svn commit: r350764 - head/sys/arm64/arm64

[Warner Losh]

On Fri, Aug 9, 2019 at 3:05 PM Konstantin Belousov <kostikbel at gmail.com>
wrote:

> On Fri, Aug 09, 2019 at 10:01:31AM -0600, Warner Losh wrote:
> > On Fri, Aug 9, 2019 at 12:57 AM Konstantin Belousov <kostikbel at gmail.com
> >
> > wrote:
> >
> > > On Thu, Aug 08, 2019 at 07:38:28PM -0600, Warner Losh wrote:
> > > > On Thu, Aug 8, 2019, 4:59 PM Gleb Smirnoff <glebius at freebsd.org>
> wrote:
> > > >
> > > > >   Hi,
> > > > >
> > > > > why do we need COMPAT_43 for arm64 at all? I can't imagine an
> > > > > application that would require this compatibility.
> > > > >
> > > > > A more general question is how far in the future are we going
> > > > > to carry COMPAT_43 for i386/amd64?
> > > > >
> > > >
> > > > COMPAT_43 is a weird option. It's a combo of both sys calls and
> kernel
> > > > behavior modifications. Before we thinned the ABIs we supported, it
> was
> > > > necessary for them as well. The biggest behavior change is around
> > > signals.
> > > > It is weird to sort out and nobody has done the deep analysis to see
> what
> > > > is truly unused and what is there for compat with Linux and other
> SysV
> > > > systems...
> > > I am not aware of any changes that COMPAT_43 provides for the signal
> > > handling semantic, except a minor adjustment for interpretation of
> > > zero-sized stack for sigaltstack(2).
> > >
> >
> > The onstack stuff was what I was thinking about, but we also have code in
> > sys_getpid() that returns the ppid in the second retval register, and
> > similar things for getuid and getgid,  It also allows ioctl numbers that
> > have IOC_IN set, but size == 0 (these would otherwise return ENOTTY). It
> > also turns on the COMPAT_OLDSOCK code which generally only kicks in when
> > compat bits are set, but in one place it allows a shorter unix domain
> > socket path length to be compatible unconditionally. The compatibility
> TTY
> > stuff, at least is under COMPAT_43TTY, but that's purely ioctl
> translation
> > code.
> I only reacted to the note about changing the signals syscalls behavior.
> But the point is valid, we should not change the syscalls ABI for new
> binaries when COMPAT_43 is enabled.  I propose the following
> https://reviews.freebsd.org/D21200


Glad I did the dumpster-diving grep then. I like your proposal for the same
reasons you stated.

WRT ioctl code for no IOC_OUT and size == 0, I believe that this is in
> fact cannot be changed. It is enabled also under COMPAT_FREEBSD4 and
> 5, and we always enable these for GENERIC. So effectively this ioctl
> permissive mode is always there.
>

Yes. I also agree. And I honestly think it's OK.


> >
> > The COMPAT_43 option indeed enables lcall 7,0 syscall entry emulation,
> > > on both i386 and amd64.  We are able to run FreeBSD 1.1.8 (i386) on
> amd64
> > > kernel in chroot this way.  Since sometimes I get bug reports about
> this
> > > stuff, there are some users of it.  I believe it is important to be
> able
> > > to run any FreeBSD binary for PR purposes, to wave the flag of
> excellent
> > > binary compatibility we offer.
> > >
> > > COMPAT_43 is there to stay as far as there are people willing to
> maintain
> > > it.  There are more than one.
> > >
> >
> > I think it's safe to retain on i386. amd64 is less clear to me, but I'd
> > lean yes.
> I believe amd64 is required since you have less and less chances to
> usefully
> run i386 kernel on modern hardware.
>

True. With your changes, enabling the option is much safer, and only drags
in a minor amount of extra code. All but the most space starved users won't
care at all about the delta in size.

Warner


> > All the other platforms I'd agree with gleb: why do we need it in
> > the kernels by default (and maybe why do we need to support it at all)?
> >
> > Warner
>