Message-ID: <20201109151750.GA10650@shinobu>
Date:   Mon, 9 Nov 2020 10:18:21 -0500
From:   William Breathitt Gray <vilhelm.gray@...il.com>
To:     Syed Nayyar Waris <syednwaris@...il.com>
Cc:     Arnd Bergmann <arnd@...nel.org>,
        Linus Walleij <linus.walleij@...aro.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        "open list:GPIO SUBSYSTEM" <linux-gpio@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Michal Simek <michal.simek@...inx.com>,
        Bartosz Golaszewski <bgolaszewski@...libre.com>,
        Andy Shevchenko <andriy.shevchenko@...ux.intel.com>,
        Linux ARM <linux-arm-kernel@...ts.infradead.org>
Subject: Re: [PATCH v12 4/4] gpio: xilinx: Utilize generic bitmap_get_value
 and _set_value

On Mon, Nov 09, 2020 at 08:18:51PM +0530, Syed Nayyar Waris wrote:
> On Mon, Nov 9, 2020 at 8:09 PM William Breathitt Gray
> <vilhelm.gray@...il.com> wrote:
> >
> > On Mon, Nov 09, 2020 at 08:41:28AM -0500, William Breathitt Gray wrote:
> > > On Mon, Nov 09, 2020 at 06:04:11PM +0530, Syed Nayyar Waris wrote:
> > > > On Sun, Nov 01, 2020 at 09:08:29PM +0100, Arnd Bergmann wrote:
> > > > > On Sun, Nov 1, 2020 at 4:00 PM William Breathitt Gray
> > > > > <vilhelm.gray@...il.com> wrote:
> > > > > >
> > > > > > On Thu, Oct 29, 2020 at 11:44:47PM +0100, Arnd Bergmann wrote:
> > > > > > > On Sun, Oct 18, 2020 at 11:44 PM Syed Nayyar Waris <syednwaris@...il.com> wrote:
> > > > > > > >
> > > > > > > > This patch reimplements the xgpio_set_multiple() function in
> > > > > > > > drivers/gpio/gpio-xilinx.c to use the new generic functions:
> > > > > > > > bitmap_get_value() and bitmap_set_value(). The code is now simpler
> > > > > > > > to read and understand. Moreover, instead of looping for each bit
> > > > > > > > in xgpio_set_multiple() function, now we can check each channel at
> > > > > > > > a time and save cycles.
> > > > > > >
> > > > > > > This now causes -Wtype-limits warnings in linux-next with gcc-10:
> > > > > >
> > > > > > Hi Arnd,
> > > > > >
> > > > > > What version of gcc-10 are you running? I'm having trouble generating
> > > > > > these warnings so I suspect I'm using a different version than you.
> > > > >
> > > > > I originally saw it with the binaries from
> > > > > https://mirrors.edge.kernel.org/pub/tools/crosstool/, but I have
> > > > > also been able to reproduce it with a minimal test case on the
> > > > > binaries from godbolt.org, see https://godbolt.org/z/Wq8q4n
> > > > >
> > > > > > Let me first verify that I understand the problem correctly. The issue
> > > > > > is the possibility of a stack smash in bitmap_set_value() when the value
> > > > > > of start + nbits is larger than the length of the map bitmap memory
> > > > > > region. This is because index (or index + 1) could be outside the range
> > > > > > of the bitmap memory region passed in as map. Is my understanding
> > > > > > correct here?
> > > > >
> > > > > Yes, that seems to be the case here.
> > > > >
> > > > > > In xgpio_set_multiple(), the variables width[0] and width[1] serve as
> > > > > > possible start and nbits values for the bitmap_set_value() calls.
> > > > > > Because width[0] and width[1] are unsigned int variables, GCC considers
> > > > > > the possibility that the value of width[0]/width[1] might exceed the
> > > > > > length of the bitmap memory region named old and thus result in a stack
> > > > > > smash.
> > > > > >
> > > > > > I don't know if invalid width values are actually possible for the
> > > > > > Xilinx gpio device, but let's err on the side of safety and assume this
> > > > > > is actually a possibility. We should verify that the combined value of
> > > > > > gpio_width[0] + gpio_width[1] does not exceed 64 bits; we can add a
> > > > > > check for this in xgpio_probe() when we grab the gpio_width values.
> > > > > >
> > > > > > However, we're still left with the GCC warnings because GCC is not smart
> > > > > > enough to know that we've already checked the boundary and width[0] and
> > > > > > width[1] are valid values. I suspect we can avoid this warning is we
> > > > > > refactor bitmap_set_value() to increment map seperately and then set it:
> > > > >
> > > > > As I understand it, part of the problem is that gcc sees the possible
> > > > > range as being constrained by the operations on 'start' and 'nbits',
> > > > > in particular the shift in BIT_WORD() that put an upper bound on
> > > > > the index, but then it sees that the upper bound is higher than the
> > > > > upper bound of the array, i.e. element zero.
> > > > >
> > > > > I added a check
> > > > >
> > > > >       if (start >= 64 || start + size >= 64) return;
> > > > >
> > > > > in the godbolt.org testcase, which does help limit the start
> > > > > index appropriately, but it is not sufficient to let the compiler
> > > > > see that the 'if (space >= nbits) ' condition is guaranteed to
> > > > > be true for all values here.
> > > > >
> > > > > > static inline void bitmap_set_value(unsigned long *map,
> > > > > >                                     unsigned long value,
> > > > > >                                     unsigned long start, unsigned long nbits)
> > > > > > {
> > > > > >         const unsigned long offset = start % BITS_PER_LONG;
> > > > > >         const unsigned long ceiling = round_up(start + 1, BITS_PER_LONG);
> > > > > >         const unsigned long space = ceiling - start;
> > > > > >
> > > > > >         map += BIT_WORD(start);
> > > > > >         value &= GENMASK(nbits - 1, 0);
> > > > > >
> > > > > >         if (space >= nbits) {
> > > > > >                 *map &= ~(GENMASK(nbits - 1, 0) << offset);
> > > > > >                 *map |= value << offset;
> > > > > >         } else {
> > > > > >                 *map &= ~BITMAP_FIRST_WORD_MASK(start);
> > > > > >                 *map |= value << offset;
> > > > > >                 map++;
> > > > > >                 *map &= ~BITMAP_LAST_WORD_MASK(start + nbits);
> > > > > >                 *map |= value >> space;
> > > > > >         }
> > > > > > }
> > > > > >
> > > > > > This avoids adding a costly conditional check inside bitmap_set_value()
> > > > > > when almost all bitmap_set_value() calls will have static arguments with
> > > > > > well-defined and obvious boundaries.
> > > > > >
> > > > > > Do you think this would be an acceptable solution to resolve your GCC
> > > > > > warnings?
> > > > >
> > > > > Unfortunately, it does not seem to make a difference, as gcc still
> > > > > knows that this compiles to the same result, and it produces the same
> > > > > warning as before (see https://godbolt.org/z/rjx34r)
> > > > >
> > > > >          Arnd
> > > >
> > > > Hi Arnd,
> > > >
> > > > Sharing a different version of bitmap_set_valuei() function. See below.
> > > >
> > > > Let me know if the below solution looks good to you and if it resolves
> > > > the above compiler warning.
> > > >
> > > >
> > > > @@ -1,5 +1,5 @@
> > > >  static inline void bitmap_set_value(unsigned long *map,
> > > > -                                    unsigned long value,
> > > > +                                    unsigned long value, const size_t length,
> > > >                                      unsigned long start, unsigned long nbits)
> > > >  {
> > > >          const size_t index = BIT_WORD(start);
> > > > @@ -7,6 +7,9 @@ static inline void bitmap_set_value(unsigned long *map,
> > > >          const unsigned long ceiling = round_up(start + 1, BITS_PER_LONG);
> > > >          const unsigned long space = ceiling - start;
> > > >
> > > > +       if (index >= length)
> > > > +               return;
> > > > +
> > > >          value &= GENMASK(nbits - 1, 0);
> > > >
> > > >          if (space >= nbits) {
> > > > @@ -15,6 +18,10 @@ static inline void bitmap_set_value(unsigned long *map,
> > > >          } else {
> > > >                  map[index + 0] &= ~BITMAP_FIRST_WORD_MASK(start);
> > > >                  map[index + 0] |= value << offset;
> > > > +
> > > > +               if (index + 1 >= length)
> > > > +                       return;
> > > > +
> > > >                  map[index + 1] &= ~BITMAP_LAST_WORD_MASK(start + nbits);
> > > >                  map[index + 1] |= value >> space;
> > > >          }
> > >
> > > One of my concerns is that we're incurring the latency two additional
> > > conditional checks just to suppress a compiler warning about a case that
> > > wouldn't occur in the actual use of bitmap_set_value(). I'm hoping
> > > there's a way for us to suppress these warnings without adding onto the
> > > latency of this function; given that bitmap_set_value() is intended to
> > > be used in loops, conditionals here could significantly increase latency
> > > in drivers.
> > >
> > > I wonder if array_index_nospec() might have the side effect of
> > > suppressing these warnings for us. For example, would this work:
> > >
> > > static inline void bitmap_set_value(unsigned long *map,
> > >                                   unsigned long value,
> > >                                   unsigned long start, unsigned long nbits)
> > > {
> > >       const unsigned long offset = start % BITS_PER_LONG;
> > >       const unsigned long ceiling = round_up(start + 1, BITS_PER_LONG);
> > >       const unsigned long space = ceiling - start;
> > >       size_t index = BIT_WORD(start);
> > >
> > >       value &= GENMASK(nbits - 1, 0);
> > >
> > >       if (space >= nbits) {
> > >               index = array_index_nospec(index, index + 1);
> > >
> > >               map[index] &= ~(GENMASK(nbits - 1, 0) << offset);
> > >               map[index] |= value << offset;
> > >       } else {
> > >               index = array_index_nospec(index, index + 2);
> > >
> > >               map[index + 0] &= ~BITMAP_FIRST_WORD_MASK(start);
> > >               map[index + 0] |= value << offset;
> > >               map[index + 1] &= ~BITMAP_LAST_WORD_MASK(start + nbits);
> > >               map[index + 1] |= value >> space;
> > >       }
> > > }
> > >
> > > Or is this going to produce the same warning because we're not using an
> > > explicit check against the map array size?
> > >
> > > William Breathitt Gray
> >
> > After testing my suggestion, it looks like the warnings are still
> > present. :-(
> >
> > Something else I've also considered is perhaps using the GCC built-in
> > function __builtin_unreachable() instead of returning. So in Syed's code
> > we would have the following instead:
> >
> > if (index + 1 >= length)
> >         __builtin_unreachable();
> >
> > This might allow GCC to optimize better and avoid the conditional check
> > all together, thus avoiding latency while also hinting enough context to
> > the compiler to suppress the warnings.
> >
> > William Breathitt Gray
> 
> I also thought of another optimization. Arnd, William, let me know
> what you think about it.
> 
> Since exceeding the array limit is a rather rare event, we can use the
> gcc extension: 'unlikely'  for the boundary checks.
> We can use it at the two places where 'index' and 'index + 1' is being
> checked against the boundary limit.
> 
> It might help optimize the code. Wouldn't it?
> 
> Syed Nayyar Waris

We probably don't need unlikely() because __builtin_unreachable() should
suffice to inform GCC that this condition will never occur -- in other
words, GCC will compile optimized code to avoid the conditional
entirely.

By the way, I think we only need the (index + 1 >= length) check; the
first index conditional check is not needed and does not affect the
warnings at all, so we might as well get rid of it.

William Breathitt Gray

Download attachment "signature.asc" of type "application/pgp-signature" (834 bytes)

Powered by blists - more mailing lists