Message-ID: <20181211115254.GC35824@arrakis.emea.arm.com>
Date:   Tue, 11 Dec 2018 11:52:55 +0000
From:   Catalin Marinas <catalin.marinas@....com>
To:     Florian Weimer <fweimer@...hat.com>
Cc:     Arnd Bergmann <arnd@...db.de>, Andy Lutomirski <luto@...nel.org>,
        "H.J. Lu" <hjl.tools@...il.com>,
        the arch/x86 maintainers <x86@...nel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Linux API <linux-api@...r.kernel.org>,
        "H. Peter Anvin" <hpa@...or.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Borislav Petkov <bp@...en8.de>, vapier@...too.org,
        Rich Felker <dalias@...c.org>, x32@...ldd.debian.org,
        Will Deacon <will.deacon@....com>,
        Linus Torvalds <torvalds@...ux-foundation.org>
Subject: Re: Can we drop upstream Linux x32 support?

On Tue, Dec 11, 2018 at 12:37:42PM +0100, Florian Weimer wrote:
> * Catalin Marinas:
> > On Tue, Dec 11, 2018 at 10:02:45AM +0100, Arnd Bergmann wrote:
> >> On Tue, Dec 11, 2018 at 6:35 AM Andy Lutomirski <luto@...nel.org> wrote:
> >> > I tried to understand what's going on.  As far as I can tell, most of
> >> > the magic is the fact that __kernel_long_t and __kernel_ulong_t are
> >> > 64-bit as seen by x32 user code.  This means that a decent number of
> >> > uapi structures are the same on x32 and x86_64.  Syscalls that only
> >> > use structures like this should route to the x86_64 entry points.  But
> >> > the implementation is still highly dubious -- in_compat_syscall() will
> >> > be *true* in such system calls,
> >> 
> >> I think the fundamental issue was that the intention had always been
> >> to use only the 64-bit entry points for system calls, but the most
> >> complex one we have -- ioctl() -- has to use the compat entry point
> >> because device drivers define their own data structures using 'long'
> >> and pointer members and they need translation, as well as
> >> matching in_compat_syscall() checks. This in turn breaks down
> >> again whenever a driver defines an ioctl command that takes
> >> a __kernel_long_t or a derived type like timespec as its argument.
> >
> > With arm64 ILP32 we tried to avoid the ioctl() problem by having
> > __kernel_long_t 32-bit, IOW mimicking the arm32 ABI (compat). The
> > biggest pain point is signals where the state is completely different
> > from arm32 (more, wider registers) and can't be dealt with by the compat
> > layer.
> 
> I would expect to approach this from the opposite direction: use 64-bit
> types in places where the 64-bit kernel interface uses 64-bit types.
> After all, not everyone who is interested in ILP32 has a companion
> 32-bit architecture which could serve as a model for the application
> ABI.

I fully agree with you that if someone wants ILP32 for a 64-bit only
architecture, they should use the 64-bit kernel interface and ensure
POSIX is adjusted.

In the arm64 context, both options were discussed with the libc
community complaining that a partial 64-bit syscall ABI breaks POSIX
while the potential users were just asking for a 32-bit ABI to run their
existing software stack on ARMv8 machines without native 32-bit support
(until they complete the migration to 64-bit).

> (If there are conflicts with POSIX, then POSIX needs to be fixed to
> support this.)

This would have been nice but no-one volunteered and, more importantly,
there was no conclusive argument that ARM ILP32 is better than LP64
(well, apart from a minority of benchmarks) and something that people
would want to migrate to. Given that the only credible case made was
about legacy code, we decided to go ahead with a (mostly) compat 32-bit
ABI.

-- 
Catalin

Powered by blists - more mailing lists