Re: [PATCH 2/2] ARM: futex: make futex_detect_cmpxchg more reliable

[Russell King - ARM Linux admin]

On Fri, Mar 08, 2019 at 09:57:45AM +0100, Ard Biesheuvel wrote:
> On Fri, 8 Mar 2019 at 00:49, Russell King - ARM Linux admin
> <linux@xxxxxxxxxxxxxxx> wrote:
> >
> > On Thu, Mar 07, 2019 at 11:39:08AM -0800, Nick Desaulniers wrote:
> > > On Thu, Mar 7, 2019 at 1:15 AM Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > > >
> > > > Passing registers containing zero as both the address (NULL pointer)
> > > > and data into cmpxchg_futex_value_locked() leads clang to assign
> > > > the same register for both inputs on ARM, which triggers a warning
> > > > explaining that this instruction has unpredictable behavior on ARMv5.
> > > >
> > > > /tmp/futex-7e740e.s: Assembler messages:
> > > > /tmp/futex-7e740e.s:12713: Warning: source register same as write-back base
> > > >
> > > > This patch was suggested by Mikael Pettersson back in 2011 (!) with gcc-4.4,
> > > > as Mikael wrote:
> > > >  "One way of fixing this is to make uaddr an input/output register, since
> > > >  "that prevents it from overlapping any other input or output."
> > > >
> > > > but then withdrawn as the warning was determined to be harmless, and it
> > > > apparently never showed up again with later gcc versions.
> > > >
> > > > Now the same problem is back when compiling with clang, and we are trying
> > > > to get clang to build the kernel without warnings, as gcc normally does.
> > > >
> > > > Cc: Mikael Pettersson <mikpe@xxxxxxxx>
> > > > Cc: Mikael Pettersson <mikpelinux@xxxxxxxxx>
> > > > Cc: Dave Martin <Dave.Martin@xxxxxxx>
> > > > Link: https://lore.kernel.org/linux-arm-kernel/20009.45690.158286.161591@xxxxxxxxxxxxxxxxxxx/
> > > > Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
> > > > ---
> > > >  arch/arm/include/asm/futex.h | 10 +++++-----
> > > >  1 file changed, 5 insertions(+), 5 deletions(-)
> > > >
> > > > diff --git a/arch/arm/include/asm/futex.h b/arch/arm/include/asm/futex.h
> > > > index 0a46676b4245..79790912974e 100644
> > > > --- a/arch/arm/include/asm/futex.h
> > > > +++ b/arch/arm/include/asm/futex.h
> > > > @@ -110,13 +110,13 @@ futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
> > > >         preempt_disable();
> > > >         __ua_flags = uaccess_save_and_enable();
> > > >         __asm__ __volatile__("@futex_atomic_cmpxchg_inatomic\n"
> > > > -       "1:     " TUSER(ldr) "  %1, [%4]\n"
> > > > -       "       teq     %1, %2\n"
> > > > +       "1:     " TUSER(ldr) "  %1, [%2]\n"
> > > > +       "       teq     %1, %3\n"
> > > >         "       it      eq      @ explicit IT needed for the 2b label\n"
> > > > -       "2:     " TUSER(streq) "        %3, [%4]\n"
> > > > +       "2:     " TUSER(streq) "        %4, [%2]\n"
> > > >         __futex_atomic_ex_table("%5")
> > > > -       : "+r" (ret), "=&r" (val)
> > > > -       : "r" (oldval), "r" (newval), "r" (uaddr), "Ir" (-EFAULT)
> > > > +       : "+&r" (ret), "=&r" (val), "+&r" (uaddr)
> > > > +       : "r" (oldval), "r" (newval), "Ir" (-EFAULT)
> > > >         : "cc", "memory");
> > > >         uaccess_restore(__ua_flags);
> > >
> > > Underspecification of constraints to extended inline assembly is a
> > > common issue exposed by other compilers (and possibly but in-effect
> > > infrequently compiler upgrades).
> > > So the reordering of the constraints means the in the assembly (notes
> > > for other reviewers):
> > > %2 -> %3
> > > %3 -> %4
> > > %4 -> %2
> > > Yep, looks good to me, thanks for finding this old patch and resending, Arnd!
> >
> > I don't see what is "underspecified" in the original constraints.
> > Please explain.
> >
> 
> I agree that that statement makes little sense.
> 
> As Russell points out in the referenced thread, there is nothing wrong
> with the generated assembly, given that the UNPREDICTABLE opcode is
> unreachable in practice. Unfortunately, we have no way to flag this
> diagnostic as a known false positive, and AFAICT, there is no reason
> we couldn't end up with the same diagnostic popping up for GCC builds
> in the future, considering that the register assignment matches the
> constraints. (We have seen somewhat similar issues where constant
> folded function clones are emitted with a constant argument that could
> never occur in reality [0])
> 
> Given the above, the only meaningful way to invoke this function is
> with different registers assigned to %3 and %4, and so tightening the
> constraints to guarantee that does not actually result in worse code
> (except maybe for the instantiations that we won't ever call in the
> first place). So I think we should fix this.
> 
> I wonder if just adding
> 
> BUG_ON(__builtin_constant_p(uaddr));
> 
> at the beginning makes any difference - this shouldn't result in any
> object code differences since the conditional will always evaluate to
> false at build time for instantiations we care about.
> 
> 
> [0] https://lore.kernel.org/lkml/9c74d635-d0d1-0893-8093-ce20b0933fc7@xxxxxxxxxx/

What I'm actually asking is:

The GCC manual says that input operands _may_ overlap output operands
since GCC assumes that input operands are consumed before output
operands are written.  This is an explicit statement.

The GCC manual does not say that input operands may overlap with each
other, and the behaviour of GCC thus far (apart from one version,
presumably caused by a bug) has been that input operands are unique.

Clang appears to be different: it allows input operands that are
registers, and contain the same constant value to be the same physical
register.

The assertion is that the constraints are under-specified.  I am
questioning that assertion.

If the constraints are under-specified, I would have expected gcc-4.4's
behaviour to have persisted, and we would've been told by gcc's
developers to fix our code.  That didn't happen, and instead gcc seems
to have been fixed.  So, my conclusion is that it is intentional that
input operands to asm() do not overlap with themselves.

It seems to me that the work-around for clang is to change every input
operand to be an output operand with a "+&r" contraint - an operand
that is both read and written by the "instruction", and that the operand
is "earlyclobber".  For something that is really only read, that seems
strange.

Also, reading GCC's manual, it would appear that "+&" is wrong.

`+'
     Means that this operand is both read and written by the
     instruction.

     When the compiler fixes up the operands to satisfy the constraints,
     it needs to know which operands are inputs to the instruction and
     which are outputs from it.  `=' identifies an output; `+'
     identifies an operand that is both input and output; all other
                                   ^^^^^^^^^^^^^^^^^^^^^
     operands are assumed to be input only.

`&'
     Means (in a particular alternative) that this operand is an
     "earlyclobber" operand, which is modified before the instruction is
     finished using the input operands.  Therefore, this operand may
     not lie in a register that is used as an input operand or as part
     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
     of any memory address.

So "+" says that this operand is an input but "&" says that it must not
be in a register that is used as an input.  That's contradictory, and I
think we can expect GCC to barf or at least end up doing strange stuff,
if not with existing versions, then with future versions.

Hence, I'm asking for clarification why it is thought that the existing
code underspecifies the asm constraints, and I'm trying to get some more
thought about what the constraints should be, in case there is a need to
use "better" constraints.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up