[PATCH] powerpc/64s: Fix possible corruption on big endian due to pgd/pud_present()

[Balbir Singh]

On Mon, Feb 18, 2019 at 11:49:18AM +1100, Michael Ellerman wrote:
> Balbir Singh <bsingharora at gmail.com> writes:
> > On Sun, Feb 17, 2019 at 07:34:20PM +1100, Michael Ellerman wrote:
> >> Balbir Singh <bsingharora at gmail.com> writes:
> >> > On Sat, Feb 16, 2019 at 08:22:12AM -0600, Segher Boessenkool wrote:
> >> >> On Sat, Feb 16, 2019 at 09:55:11PM +1100, Balbir Singh wrote:
> >> >> > On Thu, Feb 14, 2019 at 05:23:39PM +1100, Michael Ellerman wrote:
> >> >> > > In v4.20 we changed our pgd/pud_present() to check for _PAGE_PRESENT
> >> >> > > rather than just checking that the value is non-zero, e.g.:
> >> >> > > 
> >> >> > >   static inline int pgd_present(pgd_t pgd)
> >> >> > >   {
> >> >> > >  -       return !pgd_none(pgd);
> >> >> > >  +       return (pgd_raw(pgd) & cpu_to_be64(_PAGE_PRESENT));
> >> >> > >   }
> >> >> > > 
> >> >> > > Unfortunately this is broken on big endian, as the result of the
> >> >> > > bitwise && is truncated to int, which is always zero because
> >> >> 
> >> >> (Bitwise "&" of course).
> >> >> 
> >> >> > Not sure why that should happen, why is the result an int? What
> >> >> > causes the casting of pgd_t & be64 to be truncated to an int.
> >> >> 
> >> >> Yes, it's not obvious as written...  It's simply that the return type of
> >> >> pgd_present is int.  So it is truncated _after_ the bitwise and.
> >> >>
> >> >
> >> > Thanks, I am surprised the compiler does not complain about the truncation
> >> > of bits. I wonder if we are missing -Wconversion
> >> 
> >> Good luck with that :)
> >> 
> >> What I should start doing is building with it enabled and then comparing
> >> the output before and after commits to make sure we're not introducing
> >> new cases.
> >
> > Fair enough, my point was that the compiler can help out. I'll see what
> > -Wconversion finds on my local build :)
> 
> I get about 43MB of warnings here :)
>

I got about 181M with a failed build :(, but the warnings pointed to some cases
that can be a good project for cleanup

For example

1. 
static inline long regs_return_value(struct pt_regs *regs)
{
        if (is_syscall_success(regs))
                return regs->gpr[3];
        else
                return -regs->gpr[3];
}

In the case of is_syscall_success() returning false, we should ensure that
regs->gpr[3] is negative and capped within a certain limit, but it might
be an expensive check

2.
static inline void mark_hpte_slot_valid(unsigned char *hpte_slot_array,
                                        unsigned int index, unsigned int hidx)
{
        hpte_slot_array[index] = (hidx << 1) | 0x1;
}

hidx is 3 bits, but the argument is unsigned int. The caller probably does a
hidx & 0x7, but it's not clear from the code

3. hash__pmd_bad (pmd_bad) and hash__pud_bad (pud_bad) have issues similar to what was found,
but since the the page table indices are below 32, the macros are safe :)

And a few more, but I am not sure why I spent time looking at possible issues,
may be I am being stupid or overly pessimistic :)

Balbir