Re: [PATCH v10 1/4] bitops: Introduce the for_each_set_clump macro

[William Breathitt Gray]

On Sat, Oct 03, 2020 at 03:45:04PM +0300, Andy Shevchenko wrote:
> On Sat, Oct 3, 2020 at 2:37 PM Syed Nayyar Waris <syednwaris@xxxxxxxxx> wrote:
> > On Sat, Oct 3, 2020 at 2:14 PM Andy Shevchenko
> > <andy.shevchenko@xxxxxxxxx> wrote:
> > > On Sat, Oct 3, 2020 at 2:51 AM Syed Nayyar Waris <syednwaris@xxxxxxxxx> wrote:
> 
> ...
> 
> > > > +/**
> > > > + * bitmap_get_value - get a value of n-bits from the memory region
> > > > + * @map: address to the bitmap memory region
> > > > + * @start: bit offset of the n-bit value
> > > > + * @nbits: size of value in bits
> > > > + *
> > > > + * Returns value of nbits located at the @start bit offset within the @map
> > > > + * memory region.
> > > > + */
> 
> ...
> 
> > > > +               return (map[index] >> offset) & GENMASK(nbits - 1, 0);
> > >
> > > This is UB in GENMASK() when nbits == 0.
> >
> > 'nbits' actually specifies the width of clump value. Basically 'nbits'
> > denotes how-many-bits wide the clump value is.
> > 'nbits' having a value of '0' means zero-width-sized clump, meaning
> > nothing. 'nbits' can take valid values from '1' to BITS_PER_LONG.
> > The minimum value the 'nbits' can have is 1 because the smallest sized
> > clump can be 1-bit-wide. It can't be smaller than that.
> >
> > Let me know if I have misunderstood something?
> 
> It's still possible to call with an nbits parameter be equal to 0.
> If code is optimized to allow it, it should be documented that 0
> parameter is not valid and behaviour is undefined.

Documenting that 0 is not valid would be preferred because an additional
conditional check in the code could add a significant latency in a loop.
So perhaps change the documentation line to:

    @nbits: size of value in bits (must be between 1 and BITS_PER_LONG)

> 
> ...
> 
> > > > +/**
> > > > + * bitmap_set_value - set n-bit value within a memory region
> > > > + * @map: address to the bitmap memory region
> > > > + * @value: value of nbits
> > > > + * @start: bit offset of the n-bit value
> > > > + * @nbits: size of value in bits
> > > > + */
> 
> ...
> 
> > > > +       value &= GENMASK(nbits - 1, 0);
> > >
> > > This is UB when nbits == 0.
> >
> > Same as above.
> > 'nbits' actually specifies the width of clump value. Basically 'nbits'
> > denotes how-many-bits wide the clump value is.
> > 'nbits' having a value of '0' means zero-width-sized clump, meaning
> > nothing. 'nbits' can take valid values from '1' to BITS_PER_LONG.
> > The minimum value the 'nbits' can have is 1 because the smallest sized
> > clump can be 1-bit-wide. It can't be smaller than that.
> 
> Same as above.
> 
> ...
> 
> > > > +               map[index] &= ~BITMAP_FIRST_WORD_MASK(start);
> > > > +               map[index] |= value << offset;
> 
> Side note: I would prefer + 0 here and there, but it's up to you.
> 
> > > > +               map[index + 1] &= ~BITMAP_LAST_WORD_MASK(start + nbits);
> > > > +               map[index + 1] |= (value >> space);
> 
> By the way, what about this in the case of start=0, nbits > 64?
> space == 64 -> UB.
> 
> (And btw parentheses are redundant here)

I think this is the same situation as before: we should document that
nbits must be between 1 and BITS_PER_LONG.

William Breathitt Gray

> 
> > > And another LKP finding was among these lines, but I don't remember the details.
> >
> > Yes you are right. There was sparse warning reported for this.
> > sparse: shift too big (64) for type unsigned long
> > The warning was reported in patch [4/4] referring to this patch [1/4].
> >
> > Later it was clarified by the sparse-check maintainer that this
> > warning is to be ignored and no code fix is required.
> >
> > https://www.mail-archive.com/linux-kernel@xxxxxxxxxxxxxxx/msg2202377.html
> 
> Ah, okay!
> --
> With Best Regards,
> Andy Shevchenko

Attachment: signature.asc
Description: PGP signature