On Tue, Sep 03, 2024 at 06:22:42PM -0700, Nathan Chancellor wrote:
> On Tue, Sep 03, 2024 at 06:40:16PM +0300, Andy Shevchenko wrote:
> > On Tue, Sep 03, 2024 at 05:29:02PM +0200, Gergo Koteles wrote:
> > > On Tue, 2024-09-03 at 18:14 +0300, Andy Shevchenko wrote:
> > > > On Tue, Sep 03, 2024 at 05:00:51PM +0200, Gergo Koteles wrote:
> > > > > On Thu, 2024-08-29 at 19:50 +0300, Andy Shevchenko wrote:
> > > > > > First of all, it's a bit counterintuitive to have something like
> > > > > >
> > > > > > int err;
> > > > > > ...
> > > > > > scoped_guard(...)
> > > > > > err = foo(...);
> > > > > > if (err)
> > > > > > return err;
> > > > > >
> > > > > > Second, with a particular kernel configuration and compiler version in
> > > > > > one of such cases the objtool is not happy:
> > > > > >
> > > > > > ideapad-laptop.o: warning: objtool: .text.fan_mode_show: unexpected end of section
> > > > > >
> > > > > > I'm not an expert on all this, but the theory is that compiler and
> > > > > > linker in this case can't understand that 'result' variable will be
> > > > > > always initialized as long as no error has been returned. Assigning
> > > > > > 'result' to a dummy value helps with this. Note, that fixing the
> > > > > > scoped_guard() scope (as per above) does not make issue gone.
> > > > > >
> > > > > > That said, assign dummy value and make the scope_guard() clear of its scope.
> > > > > > For the sake of consistency do it in the entire file.
> > > > > >
> > > > >
> > > > > Interestingly, if I open a scope manually and use the plain guard, the
> > > > > warning disappears.
> > > >
> > > > Yes, that's what I also have, but I avoid that approach because in that case
> > > > the printing will be done inside the lock, widening the critical section for
> > > > no benefits.
> > > >
> > >
> > > This is intended to be an inner block scope within the function, it
> > > does not expand the critical section.
> >
> > I'm not sure I understand.
> >
> > scoped_guard() has a marked scope (with {} or just a line coupled with it).
> > The guard() has a scope starting at it till the end of the function. In the
> > latter case the sysfs_emit() becomes part of the critical section.
> >
> > > > > unsigned long result;
> > > > > int err;
> > > > >
> > > > > {
> > > > > guard(mutex)(&priv->vpc_mutex);
> > > > > err = read_ec_data(priv->adev->handle, VPCCMD_R_FAN,
> > > > > &result);
> > > > > if (err)
> > > > > return err;
> > > > > }
> >
> > But looking again into the code above now I got what you meant.
> > You have added a nested scope inside the function, like
> >
> > do {
> > ...
> > } while (0);
> >
> > Yes, this is strange and not what we want to have either. So I prefer to hear
> > what objtool / clang people may comment on this.
>
> So this does not appear to happen when CONFIG_KCOV is disabled with the
> configuration from the original report. I have spent some time looking
> at the disassembly but I am a little out of my element there. If I
> remember correctly, the "unexpected end of section" warning from objtool
> can appear when optimizations play fast and loose with the presence of
> potential undefined behavior (or cannot prove that there is no undefined
> behavior through inlining or analysis). In this case, I wonder if KCOV
> prevents LLVM from realizing that the for loop that scoped_guard()
> results in will run at least once, meaning that err and result would be
> potentially used uninitialized? That could explain why this change
> resolves the warning, as it ensures that no undefined behavior could
> happen regardless of whether or not the loop runs?
>
> Josh and Peter may have more insight.
Thanks for looking into this. Josh already keeps an eye on this.
--
With Best Regards,
Andy Shevchenko
