Re: [RFC PATCH 1/3] modpost: fix TO_NATIVE() with expressions and consts

[Alexander Lobakin]

From: Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx>
Date: Thu, 18 Aug 2022 16:10:21 +0200

> On Thu, Aug 18, 2022 at 04:01:53PM +0200, Alexander Lobakin wrote:
> > From: Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx>
> > Date: Thu, 18 Aug 2022 14:26:14 +0200
> > 
> > > On Thu, Aug 18, 2022 at 01:53:04PM +0200, Alexander Lobakin wrote:
> > > > Macro TO_NATIVE() directly takes a reference to its argument @x
> > > > without making an intermediate variable. This makes compilers
> > > > emit build warnings and errors if @x is an expression or a deref
> > > > of a const pointer (when target Endianness != host Endianness):
> > > > 
> > > > >> scripts/mod/modpost.h:87:18: error: lvalue required as unary '&' operand
> > > >       87 |         __endian(&(x), &(__x), sizeof(__x));                    \
> > > >          |                  ^
> > > >    scripts/mod/sympath.c:19:25: note: in expansion of macro 'TO_NATIVE'
> > > >       19 | #define t(x)            TO_NATIVE(x)
> > > >          |                         ^~~~~~~~~
> > > >    scripts/mod/sympath.c:100:31: note: in expansion of macro 't'
> > > >      100 |                 eh->e_shoff = t(h(eh->e_shoff) + off);
> > > > 
> > > > >> scripts/mod/modpost.h:87:24: warning: passing argument 2 of '__endian'
> > > > discards 'const' qualifier from pointer target type [-Wdiscarded-qualifiers]
> > > >       87 |         __endian(&(x), &(__x), sizeof(__x));                    \
> > > >          |                        ^~~~~~
> > > >    scripts/mod/sympath.c:18:25: note: in expansion of macro 'TO_NATIVE'
> > > >       18 | #define h(x)            TO_NATIVE(x)
> > > >          |                         ^~~~~~~~~
> > > >    scripts/mod/sympath.c:178:48: note: in expansion of macro 'h'
> > > >      178 |              iter < end; iter = (void *)iter + h(eh->e_shentsize)) {
> > > 
> > > How come this hasn't shown up in cross-builds today?
> > 
> > It doesn't happen with the current code.
> 
> Great, so there is no bug that you are trying to fix :)
> 
> > > > Create a temporary variable, assign @x to it and don't use @x after
> > > > that. This makes it possible to pass expressions as an argument.
> > > > Also, do a cast-away for the second argument when calling __endian()
> > > > to avoid 'discarded qualifiers' warning, as typeof() preserves
> > > > qualifiers and makes compilers think that we're passing pointer
> > > > to a const.
> > > > 
> > > > Reported-by: kernel test robot <lkp@xxxxxxxxx>
> > > > Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> > > > Cc: stable@xxxxxxxxxxxxxxx # 4.9+
> > > 
> > > Where are these build warnings showing up at that we don't see them
> > > today, yet this is needed to go back to all stable trees?
> > 
> > I thought all fixes should go to the applicable stable trees, am I
> > wrong? If so, I'll drop the tag in the next spin.
> 
> But this isn't fixing a bug in the code today that anyone can hit, so
> why would you mark it as such?

So do you mean that a fix is a fix not when it makes some wrong code
work properly, but only when there's a certain bug report and this
fix seems to resolve it?
I.e, if there are no ways to reach some code in which 2 + 2 == 5,
there is no bug? A loaded shotgun can't be considered loaded unless
someone shots his leg?

I mean, I understand the rule "don't touch if it works", but dunno,
I don't feel it's: 1) completely justified; 2) always followed in
the current stable trees.
But I'm not a -stable maintainer :)

> 
> > I remember we had such discussion already regarding fixing stuff in
> > modpost, which can happen only with never mainlained GCC LTO or with
> > the in-dev code. At the end that fix made it into the stables IIRC.
> 
> I don't remember taking fixes for out-of-tree LTO stuff, but I shouldn't
> have :)

This: [0]

There is no way to repro it on the stable kernels, but it's here
backported :)

> 
> thanks,
> 
> greg k-h

[0] https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?h=linux-4.19.y&id=03bd6eaab3e1cbd4e5060b36a67000165f6e0482

Thanks,
Olek