Re: [RFC][PATCH 5/7] x86: convert arch_futex_atomic_op_inuser() to user_access_begin/user_access_end()

[Al Viro]

On Mon, Mar 23, 2020 at 12:06:39PM -0700, Linus Torvalds wrote:
> On Mon, Mar 23, 2020 at 11:53 AM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > From: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
> >
> > Lift stac/clac pairs from __futex_atomic_op{1,2} into arch_futex_atomic_op_inuser(),
> > fold them with access_ok() in there.
> 
> So this is a deep internal macro and already has the double
> underscore, but I'm inclined to say "add the unsafe here too" for
> those __futex_atomic_opX() macros that are now called inside that
> user_access_begin/end region.
> 
> And wouldn't it be lovely to get rid of the error return thing, and
> pass in a label instead, the way "usafe_get/put_user()" works too?
> That might be a separate patch from the "reorg" thing, though.

Umm...
#define __futex_atomic_op1(insn, ret, oldval, uaddr, oparg)     \
        asm volatile("1:\t" insn "\n"                           \
                     "2:\n"                                     \
                     "\t.section .fixup,\"ax\"\n"               \
                     "3:\tmov\t%3, %1\n"                        \
                     "\tjmp\t2b\n"                              \
                     "\t.previous\n"                            \
                     _ASM_EXTABLE_UA(1b, 3b)                    \
                     : "=r" (oldval), "=r" (ret), "+m" (*uaddr) \
                     : "i" (-EFAULT), "0" (oparg), "1" (0))

OK, ret wouldn't be in the list of outputs that way and
*uaddr could become an input (we only care about the address,
same as for put_user), but oldval is a genuine output -
insn is xchgl or lock xaddl, and we obviously care about the
old value fetched by it.  The same goes for
#define __futex_atomic_op2(insn, ret, oldval, uaddr, oparg)     \
        asm volatile("1:\tmovl  %2, %0\n"                       \
                     "\tmovl\t%0, %3\n"                         \
                     "\t" insn "\n"                             \
                     "2:\t" LOCK_PREFIX "cmpxchgl %3, %2\n"     \
                     "\tjnz\t1b\n"                              \
                     "3:\n"                                     \
                     "\t.section .fixup,\"ax\"\n"               \
                     "4:\tmov\t%5, %1\n"                        \
                     "\tjmp\t3b\n"                              \
                     "\t.previous\n"                            \
                     _ASM_EXTABLE_UA(1b, 4b)                    \
                     _ASM_EXTABLE_UA(2b, 4b)                    \
                     : "=&a" (oldval), "=&r" (ret),             \
                       "+m" (*uaddr), "=&r" (tem)               \
                     : "r" (oparg), "i" (-EFAULT), "1" (0))
- oldval is calculated by that thing and arch_futex_atomic_op_inuser()
ends up storing it in *oval.  And moving that assignment into
the inline asm will simply put *oval into the output list,
won't it?

How would you work around that?