Re: [PATCH] sched/cputime: silence a -Wunused-function warning

[Nick Desaulniers <ndesaulniers@google.com>]

On Fri, Mar 6, 2020 at 10:39 AM Joe Perches <joe@perches.com> wrote:
>
> On Fri, 2020-03-06 at 09:25 -0800, Joe Perches wrote:
> > On Fri, 2020-03-06 at 09:13 -0800, Nick Desaulniers wrote:
> > > On Fri, Mar 6, 2020 at 7:42 AM Qian Cai <cai@lca.pw> wrote:
> > > > account_other_time() is only used when CONFIG_IRQ_TIME_ACCOUNTING=y (in
> > > > irqtime_account_process_tick()) or CONFIG_VIRT_CPU_ACCOUNTING_GEN=y (in
> > > > get_vtime_delta()). When both are off, it will generate a compilation
> > > > warning from Clang,
> > > >
> > > > kernel/sched/cputime.c:255:19: warning: unused function
> > > > 'account_other_time' [-Wunused-function]
> > > > static inline u64 account_other_time(u64 max)
> > > >
> > > > Rather than wrapping around this function with a macro expression,
> > > >
> > > >  if defined(CONFIG_IRQ_TIME_ACCOUNTING) || \
> > > >     defined(CONFIG_VIRT_CPU_ACCOUNTING_GEN)
> > > >
> > > > just use __maybe_unused for this small function which seems like a good
> > > > trade-off.
> > >
> > > Generally, I'm not a fan of __maybe_unused.  It is a tool in the
> > > toolbox for solving this issue, but it's my least favorite.  Should
> > > the call sites be eliminated, this may mask an unused and entirely
> > > dead function from being removed.  Preprocessor guards based on config
> > > give more context into *why* a particular function may be unused.
> > >
> > > So let's take a look at the call sites of account_other_time().  Looks
> > > like irqtime_account_process_tick() (guarded by
> > > CONFIG_IRQ_TIME_ACCOUNTING) and get_vtime_delta() (guarded by
> > > CONFIG_VIRT_CPU_ACCOUNTING_GEN).  If it were one config guard, then I
> > > would prefer to move the definition to be within the same guard, just
> > > before the function definition that calls it, but we have a more
> > > complicated case here.
> > >
> > > The next thing I'd check to see is if there's a dependency between
> > > configs.  In this case, both CONFIG_IRQ_TIME_ACCOUNTING and
> > > CONFIG_VIRT_CPU_ACCOUNTING_GEN are defined in init/Kconfig.  In this
> > > case there's also no dependency between configs, so specifying one
> > > doesn't imply the other; so guarding on one of the two configs is also
> > > not an option.
> > >
> > > So, as much as I'm not a fan of __maybe_unused, it is indeed the
> > > simplest option.  Though, I'd still prefer the ifdef you describe
> > > instead, maybe the maintainers can provide guidance/preference?
> >
> > Another option might be to move static inline functions
> > where possible to an #include file (like sched.h) but the
> > same possible dead function issue would still exist.
>
> Turns out there are hundreds of unused static inline
> functions in kernel .h files.
>
> A trivial script to find some of them (with likely
> false positives as some might actually be used via ##
> token pasting mechanisms).
>
> (and there's almost certainly a better way to do this
>  too as it takes a _very_ long time to run)
>
> $ grep-2.5.4 -rP --include=*.h '^[ \t]*static\s+inline\s+(?:\w+\s+){1,3}\w+[ \t]*\(' * | \
>   grep -P -oh '\w+\s*\(' | \
>   sort | uniq -c | sort -n | grep -P '^\s+1\b' | \
>   sed -r -e 's/^\s+1\s+//' -e 's/\(//' | \
>   while read line ; do \
>     echo -n "$line: " ; git grep -w $line | wc -l ; \
>   done | \
>   grep ": 1$"

Please start sending patches to remove them and I'll review.  If this
is a good amount of work, I have newbies that are looking to
contribute and can help.
-- 
Thanks,
~Nick Desaulniers