Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
References: <cbecab05-9e95-5dec-ef81-499617c153a6@suse.com>
 <08B92B44-CCA9-4B83-B9CC-F1601D44B73F@amacapital.net> <0053f606-f4f7-3951-f40b-b7bd08703590@suse.com>
 <CALCETrWHNunMzP1xHmOhvHG20_baoeXhNbCcEJgCgm5xzGM5Tw@mail.gmail.com>
 <ed9d8df6-0fe7-ca15-bab2-4d9cbbfe62f0@suse.com> <CALCETrXoK+6gNn=3_yZdkHScd=N-a2f_VPC-svkFfHVsiVusVw@mail.gmail.com>
 <62d9f87d-de55-3fb4-664d-d24897f4dd9b@suse.com> <CAMzpN2gBrXWwUrDUQQAatutg6xxgxOunVrse6gs2C=EYKH11JA@mail.gmail.com>
In-Reply-To: <CAMzpN2gBrXWwUrDUQQAatutg6xxgxOunVrse6gs2C=EYKH11JA@mail.gmail.com>
From:   Andy Lutomirski <luto@kernel.org>
Date:   Wed, 18 Dec 2019 18:39:09 -0800
Message-ID: <CALCETrXDhPU54L8XmXanXK4pTPKykCVx44xahrvoZ8K_ZdTMRw@mail.gmail.com>
Subject: Re: [PATCH] x86-64/entry: add instruction suffix to SYSRET
To:     Brian Gerst <brgerst@gmail.com>
Cc:     Jan Beulich <jbeulich@suse.com>, Andy Lutomirski <luto@kernel.org>,
        "the arch/x86 maintainers" <x86@kernel.org>,
        lkml <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Mon, Dec 16, 2019 at 7:23 AM Brian Gerst <brgerst@gmail.com> wrote:
>
> On Mon, Dec 16, 2019 at 5:12 AM Jan Beulich <jbeulich@suse.com> wrote:
> >
> > On 13.12.2019 18:49, Andy Lutomirski wrote:
> > > On Fri, Dec 13, 2019 at 1:55 AM Jan Beulich <jbeulich@suse.com> wrote=
:
> > >>
> > >> On 12.12.2019 22:43, Andy Lutomirski wrote:
> > >>> On Tue, Dec 10, 2019 at 7:40 AM Jan Beulich <jbeulich@suse.com> wro=
te:
> > >>>>
> > >>>> On 10.12.2019 16:29, Andy Lutomirski wrote:
> > >>>>>> On Dec 10, 2019, at 2:48 AM, Jan Beulich <JBeulich@suse.com> wro=
te:
> > >>>>>>
> > >>>>>> =EF=BB=BFOmitting suffixes from instructions in AT&T mode is bad=
 practice when
> > >>>>>> operand size cannot be determined by the assembler from register
> > >>>>>> operands, and is likely going to be warned about by upstream gas=
 in the
> > >>>>>> future. Add the missing suffix here.
> > >>>>>>
> > >>>>>> Signed-off-by: Jan Beulich <jbeulich@suse.com>
> > >>>>>>
> > >>>>>> --- a/arch/x86/entry/entry_64.S
> > >>>>>> +++ b/arch/x86/entry/entry_64.S
> > >>>>>> @@ -1728,7 +1728,7 @@ END(nmi)
> > >>>>>> SYM_CODE_START(ignore_sysret)
> > >>>>>>    UNWIND_HINT_EMPTY
> > >>>>>>    mov    $-ENOSYS, %eax
> > >>>>>> -    sysret
> > >>>>>> +    sysretl
> > >>>>>
> > >>>>> Isn=E2=80=99t the default sysretq?  sysretl looks more correct, b=
ut that suggests
> > >>>>> that your changelog is wrong.
> > >>>>
> > >>>> No, this is different from ret, and more like iret and lret.
> > >>>>
> > >>>>> Is this code even reachable?
> > >>>>
> > >>>> Yes afaict, supported by the comment ahead of the symbol. syscall_=
init()
> > >>>> puts its address into MSR_CSTAR when !IA32_EMULATION.
> > >>>>
> > >>>
> > >>> What I meant was: can a program actually get itself into 32-bit mod=
e
> > >>> to execute a 32-bit SYSCALL instruction?
> > >>
> > >> Why not? It can set up a 32-bit code segment descriptor, far-branch
> > >> into it, and then execute SYSCALL. I can't see anything preventing
> > >> this in the logic involved in descriptor adjustment system calls. In
> > >> fact it looks to be at least partly the opposite - fill_ldt()
> > >> disallows creation of 64-bit code segments (oddly enough
> > >> fill_user_desc() then still copies the bit back, despite there
> > >> apparently being no way for it to get set).
> > >
> > > Do we allow creation of 32-bit code segments on !IA32_EMULATION
> > > kernels?
> >
> > As per above - I think so.
> >
> > >  I think we shouldn't, but I'm not really sure.
> >
> > It may be a little exotic, but I can't see any reason to disallow
> > a 64-bit process to switch to compatibility mode temporarily. One
> > contrived use case could be to be able to invoke INTO or BOUND.
>
> I think it should be kept intact for future use by WINE.  WINE is
> currently set up so that 32/16-bit Windows emulation needs a 32-bit
> build against 32-bit Linux libraries, using the kernel compat layer.
> With many distributions wanting to drop 32-bit support this has been a
> big sticking point.  If WINE could be modified so that the core is
> always built as 64-bit with 32-bit compatibility handled entirely in
> userspace, that would remove its dependency on 32-bit Linux libraries
> and thus wouldn't require IA32_EMULATION.

I just read an article about WINE on Mac OS doing more or less this.
I'm wondering if we should wire up set_thread_area() on x86_64 for
uses like this.  It even has a syscall number already -- it's just not
wired up.

Anyway, this isn't particularly high priority, IMO -- I don't think
many people set IA32_EMULATION=3Dn.  What we really ought to do is to
get rid of the special ignore_sysret path and instead go through the
normal syscall path but just do:

if (!IS_ENABLED(IA32_EMULATION))
  return -ENOSYS;

or equivalent.  The last thing we need is a whole special asm path
that essentially no one uses.