Re: linux-next: Tree for Aug 20 (Wno-alloc-size-larger-than)

[Randy Dunlap]

On 8/25/21 8:54 PM, Kees Cook wrote:
> On Wed, Aug 25, 2021 at 10:49:19AM -0700, Randy Dunlap wrote:
> > On 8/25/21 10:04 AM, Kees Cook wrote:
> > > On Tue, Aug 24, 2021 at 11:58:59AM +1000, Stephen Rothwell wrote:
> > > > On Mon, 23 Aug 2021 18:24:44 -0700 Randy Dunlap <rdunlap@xxxxxxxxxxxxx> wrote:
> > > > > This is just weird. What I am seeing is that for every source file
> > > > > where gcc emits a warning: it then follows that up with this
> > > > > > > cc1: warning: unrecognized command line option '-Wno-alloc-size-larger-than'
> > > >
> > > > I see the same, as well as:
> > > >
> > > > <stdin>:1515:2: warning: #warning syscall clone3 not implemented [-Wcpp]
> > > > cc1: warning: unrecognized command line option '-Wno-alloc-size-larger-than'
> > > >
> > > > But only on my gcc 7.3.1 builds (the rest are gcc 10).
> > > >
> > > > > Smells like a gcc bug to me.
> > > >
> > > > Yes
> > > >
> > > > Also noted here: https://github.com/DynamoRIO/drmemory/issues/2099 (second comment)
> > >
> > > Okay, I think this work-around should work. I've been able to reproduce
> > > the weird conditions, and this seems to behave correctly. Andrew, can
> > > you fixup the fixup with this?
> > >
> > >
> > > diff --git a/Makefile b/Makefile
> > > index 26640899e7ca..c1842014a5de 100644
> > > --- a/Makefile
> > > +++ b/Makefile
> > > @@ -1094,8 +1094,13 @@ endif
> > >    ifdef CONFIG_CC_IS_GCC
> > >    # The allocators already balk at large sizes, so silence the compiler
> > > -# warnings for bounds checks involving those possible values.
> > > -KBUILD_CFLAGS += $(call cc-option, -Wno-alloc-size-larger-than)
> > > +# warnings for bounds checks involving those possible values. While
> > > +# -Wno-alloc-size-larger-than would normally be used here, some versions
> > > +# of gcc (<9.1) weirdly don't handle the option correctly when _other_
> > > +# warnings are produced (?!), so instead use SIZE_MAX to effectively
> > > +# disable it.
> > > +# https://lore.kernel.org/lkml/20210824115859.187f272f@xxxxxxxxxxxxxxxx
> > > +KBUILD_CFLAGS += $(call cc-option, -Walloc-size-larger-than=SIZE_MAX)
> > >    endif
> > >    # disable invalid "can't wrap" optimizations for signed / pointers
> >
> > Hi Kees,
> >
> > I get a lot of these:
> >
> > ../include/linux/slab.h: In function ‘keyctl_instantiate_key_common’:
> > cc1: warning: invalid argument ‘SIZE_MAX’ to ‘-Walloc-size-larger-than=’
>
> O_o
>
> I love how the documentation on this option is consistently wrong. :)
>
> I haven't been able to exactly reproduce this error on godbolt.org, but
> I got close with trunk GCC:
> gcc: error: argument to '-Walloc-size-larger-than=' should be a non-negative integer optionally followed by a size unit
>
> Even though stdint.h is included. :(
>
> Okay. How about _this_ fix?
>
> diff --git a/Makefile b/Makefile
> index efa9bd36b158..141a851930e6 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -1096,8 +1096,17 @@ endif
>   
>   ifdef CONFIG_CC_IS_GCC
>   # The allocators already balk at large sizes, so silence the compiler
> -# warnings for bounds checks involving those possible values.
> -KBUILD_CFLAGS += $(call cc-option, -Wno-alloc-size-larger-than)
> +# warnings for bounds checks involving those possible values. While
> +# -Wno-alloc-size-larger-than would normally be used here, earlier versions
> +# of gcc (<9.1) weirdly don't handle the option correctly when _other_
> +# warnings are produced (?!). Using -Walloc-size-larger-than=SIZE_MAX
> +# doesn't work (as it is documented to), silently resolving to "0" prior to
> +# version 9.1 (and producing an error more recently). Numeric values larger
> +# than PTRDIFF_MAX also don't work prior to version 9.1, which are silently
> +# ignored, continuing to default to PTRDIFF_MAX. So, left with no other
> +# choice, we must perform a versioned check to disable this warning.
> +# https://lore.kernel.org/lkml/20210824115859.187f272f@xxxxxxxxxxxxxxxx
> +KBUILD_CFLAGS += $(call cc-ifversion, -ge, 0901, -Wno-alloc-size-larger-than)
>   endif
>   
>   # disable invalid "can't wrap" optimizations for signed / pointers

Yes, this works for me. Thanks.

Tested-by: Randy Dunlap <rdunlap@xxxxxxxxxxxxx>

--
~Randy