Re: [PATCH 01/17] kbuild: Disable -Waddress-of-packed-member for gcc 9

[Arnd Bergmann]

On Fri, Mar 22, 2019 at 5:40 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Fri, Mar 22, 2019 at 02:58:51PM +0100, Arnd Bergmann wrote:
> > On Thu, Mar 21, 2019 at 11:00 PM Andi Kleen <andi@xxxxxxxxxxxxxx> wrote:
> > >
> > > From: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> > >
> > > This warning is very noisy in a default build with gcc 9.
> > > Move it into W=2 only.
> > >
> > > Cc: arnd@xxxxxxxx
> > > Cc: Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx>
> > > Signed-off-by: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> >
> > I think W=2 is too aggressive. On many architectures, this finds
> > real bugs and the false positives tend to be trivial to fix
> > (by removing the bogus __packed annotation), which improves
> > the generated code in the process.
> >
> > Moving it to W=1 for the moment is probably fine, but ideally
> > I think we should fix the kernel to behave according to the
> > C standard.
>
> Lol... we're actively moving away from the C standard on many places.
>
> Why does the silly compiler think it is a problem to take the address of
> a member of a packed structure? That sounds like something that's
> perfectly fine and useful.

The problem is casting a pointer of an unaligned variable to one
of higher alignment and passing that to a function that does not
expect unaligned data.

On CPUs that don't have unaligned load/store instructions, this
causes an alignment fault, which will have to be fixed up in
software by doing a bytewise access or crash the system.

        Arnd