Re: [PATCH v2 21/29] nios2: Futex operations

[Arnd Bergmann <arnd@arndb.de>]

On Thursday 17 July 2014 18:55:49 Ley Foon Tan wrote:
> On Tue, Jul 15, 2014 at 6:03 PM, Thomas Gleixner <tglx@linutronix.de> wrote:
> > On Tue, 15 Jul 2014, Ley Foon Tan wrote:
> >  +static inline int atomic_futex_op_xchg_set(int oparg, u32 __user *uaddr,
> >> +                                        int *oldval)
> >> +{
> >> +     unsigned long flags;
> >> +     int ret;
> >> +
> >> +     local_irq_save(flags);
> >> +
> >> +     ret = get_user(*oldval, uaddr);
> >> +     if (!ret)
> >> +             ret = put_user(oparg, uaddr);
> >
> > This is wrong as it gets. get_user() might fault and sleep.
> >
> > You need a proper implementation, which handles fault exceptions.
> I have checked that we use nios2 specific get_user() in [1]. This
> function will not sleep and it handles fault exception.
> I think this should be fine.

The get_user/put_user functions really need to be annotated might_fault(),
because that's what they do.

The whole point of get_user() is to access an unchecked user space
pointer, which 	can do a number of things based on what the pointer
points to:

- access a user space variable that resides in memory
- access a kernel pointer and fail because of the access_ok()
  check
- access a user space pointer that is not mapped and return
  through the __ex_table fixup.
- access a user space pointer that has a valid VMA but not PTE,
  causing a page fault to be resolved.

It's the last case that breaks here.

	Arnd